Use fixed point math for the inner loop of ZoomSurface(). This gives a huge speedup on PowerPC.
Note that 16777216=2^24 is 1.0 in U(8,24) fixed point and 8388608=2^23 is 0.5. Also, single precision IEEE floating point has 23 bits of mantissa and U(8,24) has 24 bits of fraction so no precision is lost.
This commit is contained in:
@@ -16,7 +16,7 @@ using namespace std;
|
||||
* lines.)
|
||||
*/
|
||||
|
||||
static void InitVectors( vector<int> &s0, vector<int> &s1, vector<float> &percent, int src, int dst )
|
||||
static void InitVectors( vector<int> &s0, vector<int> &s1, vector<uint32_t> &percent, int src, int dst )
|
||||
{
|
||||
if( src >= dst )
|
||||
{
|
||||
@@ -28,12 +28,12 @@ static void InitVectors( vector<int> &s0, vector<int> &s1, vector<float> &percen
|
||||
* going 512->256, then dst[0] should come from the pixels from 0..1 and
|
||||
* 1..2, so sax[0] is 1. sx is the total number of pixels, so sx/2 is the
|
||||
* distance from the start of the sample to its center. */
|
||||
const float sax = sx*x + sx/2;
|
||||
const float sax = sx*x + sx/2.0f;
|
||||
|
||||
/* sx/2 is the distance from the start of the sample to the center;
|
||||
* sx/4 is the distance from the center of the sample to the center of
|
||||
* either pixel. */
|
||||
const float xstep = sx/4;
|
||||
const float xstep = sx/4.0f;
|
||||
|
||||
/* source x coordinates of left and right pixels to sample */
|
||||
s0.push_back(int(sax-xstep));
|
||||
@@ -43,7 +43,7 @@ static void InitVectors( vector<int> &s0, vector<int> &s1, vector<float> &percen
|
||||
{
|
||||
/* If the sampled pixels happen to be the same, the distance
|
||||
* will be 0. Avoid division by zero. */
|
||||
percent.push_back( 1.f );
|
||||
percent.push_back( 1<<24 );
|
||||
} else {
|
||||
const int xdist = s1[x] - s0[x];
|
||||
|
||||
@@ -52,8 +52,8 @@ static void InitVectors( vector<int> &s0, vector<int> &s1, vector<float> &percen
|
||||
|
||||
/* sax is somewhere between the centers of both sampled
|
||||
* pixels; find the percentage: */
|
||||
const float p = (sax - fleft) / xdist;
|
||||
percent.push_back(1-p);
|
||||
const float p = (1.0f - (sax - fleft) / xdist) * 16777216.0f;
|
||||
percent.push_back( uint32_t(p) );
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -77,8 +77,8 @@ static void InitVectors( vector<int> &s0, vector<int> &s1, vector<float> &percen
|
||||
s0.push_back( clamp(int(sax), 0, src-1));
|
||||
s1.push_back( clamp(int(sax+1), 0, src-1) );
|
||||
|
||||
const float p = 1 - (sax - floorf(sax));
|
||||
percent.push_back( p );
|
||||
const float p = (1.0f - (sax - floorf(sax))) * 16777216.0f;
|
||||
percent.push_back( uint32_t(p) );
|
||||
}
|
||||
}
|
||||
}
|
||||
@@ -88,7 +88,7 @@ static void ZoomSurface( const RageSurface * src, RageSurface * dst )
|
||||
/* For each destination coordinate, two source rows, two source columns
|
||||
* and the percentage of the first row and first column: */
|
||||
vector<int> esx0, esx1, esy0, esy1;
|
||||
vector<float> ex0, ey0;
|
||||
vector<uint32_t> ex0, ey0;
|
||||
|
||||
InitVectors( esx0, esx1, ex0, src->w, dst->w );
|
||||
InitVectors( esy0, esy1, ey0, src->h, dst->h );
|
||||
@@ -114,13 +114,14 @@ static void ZoomSurface( const RageSurface * src, RageSurface * dst )
|
||||
uint8_t color[4];
|
||||
for( int c = 0; c < 4; ++c )
|
||||
{
|
||||
float x0, x1;
|
||||
x0 = c00[c] * ex0[x];
|
||||
x0 += c01[c] * (1-ex0[x]);
|
||||
x1 = c10[c] * ex0[x];
|
||||
x1 += c11[c] * (1-ex0[x]);
|
||||
|
||||
const float res = (x0 * ey0[y]) + (x1 * (1-ey0[y])) + 0.5f;
|
||||
uint32_t x0 = uint32_t(c00[c]) * ex0[x];
|
||||
x0 += uint32_t(c01[c]) * (16777216 - ex0[x]);
|
||||
x0 >>= 24;
|
||||
uint32_t x1 = uint32_t(c10[c]) * ex0[x];
|
||||
x1 += uint32_t(c11[c]) * (16777216 - ex0[x]);
|
||||
x1 >>= 24;
|
||||
|
||||
const uint32_t res = ((x0 * ey0[y]) + (x1 * (16777216-ey0[y])) + 8388608) >> 24;
|
||||
color[c] = uint8_t(res);
|
||||
}
|
||||
*dp = *(uint32_t *) color;
|
||||
|
||||
Reference in New Issue
Block a user