/* Generate code for a section of the mask. first is the index we start
* at, we set last to the index of the last one we use before we run
* out of intermediates / constants / parameters / sources or mask
* coefficients.
*
* 0 for success, -1 on error.
*/
static int
vips_reducev_compile_section( VipsReducev *reducev, Pass *pass, gboolean first )
{
VipsVector *v;
int i;
#ifdef DEBUG_COMPILE
printf( "starting pass %d\n", pass->first );
#endif /*DEBUG_COMPILE*/
pass->vector = v = vips_vector_new( "reducev", 1 );
/* We have two destinations: the final output image (8-bit) and the
* intermediate buffer if this is not the final pass (16-bit).
*/
pass->d2 = vips_vector_destination( v, "d2", 2 );
/* "r" is the array of sums from the previous pass (if any).
*/
pass->r = vips_vector_source_name( v, "r", 2 );
/* The value we fetch from the image, the accumulated sum.
*/
TEMP( "value", 2 );
TEMP( "sum", 2 );
/* Init the sum. If this is the first pass, it's a constant. If this
* is a later pass, we have to init the sum from the result
* of the previous pass.
*/
if( first ) {
char c0[256];
CONST( c0, 0, 2 );
ASM2( "loadpw", "sum", c0 );
}
else
ASM2( "loadw", "sum", "r" );
for( i = pass->first; i < reducev->n_point; i++ ) {
char source[256];
char coeff[256];
SCANLINE( source, i, 1 );
/* This mask coefficient.
*/
vips_snprintf( coeff, 256, "p%d", i );
pass->p[pass->n_param] = PARAM( coeff, 2 );
pass->n_param += 1;
if( pass->n_param >= MAX_PARAM )
return( -1 );
/* Mask coefficients are 2.6 bits fixed point. We need to hold
* about -0.5 to 1.0, so -2 to +1.999 is as close as we can
* get.
*
* We need a signed multiply, so the image pixel needs to
* become a signed 16-bit value. We know only the bottom 8 bits
* of the image and coefficient are interesting, so we can take
* the bottom bits of a 16x16->32 multiply.
*
* We accumulate the signed 16-bit result in sum.
*/
ASM2( "convubw", "value", source );
ASM3( "mullw", "value", "value", coeff );
ASM3( "addssw", "sum", "sum", "value" );
/* We've used this coeff.
*/
pass->last = i;
if( vips_vector_full( v ) )
break;
/* orc 0.4.24 and earlier hate more than about five lines at
* once :(
*/
if( i - pass->first > 3 )
break;
}
/* If this is the end of the mask, we write the 8-bit result to the
* image, otherwise write the 16-bit intermediate to our temp buffer.
*/
if( pass->last >= reducev->n_point - 1 ) {
char c32[256];
char c6[256];
char c0[256];
char c255[256];
CONST( c32, 32, 2 );
ASM3( "addw", "sum", "sum", c32 );
CONST( c6, 6, 2 );
ASM3( "shrsw", "sum", "sum", c6 );
/* You'd think "convsuswb", convert signed 16-bit to unsigned
* 8-bit with saturation, would be quicker, but it's a lot
* slower.
*/
CONST( c0, 0, 2 );
ASM3( "maxsw", "sum", c0, "sum" );
CONST( c255, 255, 2 );
ASM3( "minsw", "sum", c255, "sum" );
ASM2( "convwb", "d1", "sum" );
}
else
ASM2( "copyw", "d2", "sum" );
if( !vips_vector_compile( v ) )
return( -1 );
#ifdef DEBUG_COMPILE
printf( "done coeffs %d to %d\n", pass->first, pass->last );
vips_vector_print( v );
#endif /*DEBUG_COMPILE*/
return( 0 );
}
/* Generate code for a section of the mask. first is the index we start
* at, we set last to the index of the last one we use before we run
* out of intermediates / constants / parameters / sources or mask
* coefficients.
*
* 0 for success, -1 on error.
*/
static int
pass_compile_section( Pass *pass, Morph *morph, gboolean first_pass )
{
INTMASK *mask = morph->mask;
const int n_mask = mask->xsize * mask->ysize;
VipsVector *v;
char offset[256];
char source[256];
char zero[256];
char one[256];
int i;
pass->vector = v = vips_vector_new( "morph", 1 );
/* The value we fetch from the image, the accumulated sum.
*/
TEMP( "value", 1 );
TEMP( "sum", 1 );
CONST( zero, 0, 1 );
CONST( one, 255, 1 );
/* Init the sum. If this is the first pass, it's a constant. If this
* is a later pass, we have to init the sum from the result
* of the previous pass.
*/
if( first_pass ) {
if( morph->op == DILATE )
ASM2( "copyb", "sum", zero );
else
ASM2( "copyb", "sum", one );
}
else {
/* "r" is the result of the previous pass.
*/
pass->r = vips_vector_source_name( v, "r", 1 );
ASM2( "loadb", "sum", "r" );
}
for( i = pass->first; i < n_mask; i++ ) {
int x = i % mask->xsize;
int y = i / mask->xsize;
/* Exclude don't-care elements.
*/
if( mask->coeff[i] == 128 )
continue;
/* The source. sl0 is the first scanline in the mask.
*/
SCANLINE( source, y, 1 );
/* The offset, only for non-first-columns though.
*/
if( x > 0 ) {
CONST( offset, morph->in->Bands * x, 1 );
ASM3( "loadoffb", "value", source, offset );
}
else
ASM2( "loadb", "value", source );
/* Join to our sum. If the mask element is zero, we have to
* add an extra negate.
*/
if( morph->op == DILATE ) {
if( !mask->coeff[i] )
ASM3( "xorb", "value", "value", one );
ASM3( "orb", "sum", "sum", "value" );
}
else {
if( !mask->coeff[i] )
ASM3( "andnb", "sum", "sum", "value" );
else
ASM3( "andb", "sum", "sum", "value" );
}
if( vips_vector_full( v ) )
break;
}
pass->last = i;
ASM2( "copyb", "d1", "sum" );
if( !vips_vector_compile( v ) )
return( -1 );
#ifdef DEBUG
printf( "done matrix coeffs %d to %d\n", pass->first, pass->last );
vips_vector_print( v );
#endif /*DEBUG*/
return( 0 );
}