[sldev] OpenJPEG -> dwt -> alloc -> optimization

[Dzonatas]

Here is a patch I started to work with today.  I ripped this out of my 
vectorized version. This patch just creates two new functions to 
allocate memory to the maximum computed level. It's just single step in 
optimization of what I have.

I'll post the encode version as a separate patch if this fly's.

-------------- next part --------------
Index: libopenjpeg/dwt.c
===================================================================
--- libopenjpeg/dwt.c	(revision 369)
+++ libopenjpeg/dwt.c	(working copy)
@@ -95,6 +95,14 @@
 FIXME : comment ???
 */
 static void dwt_encode_stepsize(int stepsize, int numbps, opj_stepsize_t *bandno_stepsize);
+/*
+Allocate memory for maximum computed resolution level in decode
+*/
+static int* dwt_decode_alloc(opj_tcd_tilecomp_t * tilec, int stop);
+/*
+Free memory used for maximum computed resolution level in decode
+*/
+static void dwt_decode_free(int* m);
 
 /*@}*/
 
@@ -400,6 +408,7 @@
 	w = tilec->x1-tilec->x0;
 	l = tilec->numresolutions-1;
 	a = tilec->data;
+	bj = dwt_decode_alloc(tilec, stop);
 	
 	for (i = l - 1; i >= stop; i--) {
 		int rw;			/* width of the resolution level computed                                                           */
@@ -420,7 +429,6 @@
 		
 		sn = rw1;
 		dn = rw - rw1;
-		bj = (int*)opj_malloc(rw * sizeof(int));
 		for (j = 0; j < rh; j++) {
 			aj = a + j*w;
 			dwt_interleave_h(aj, bj, dn, sn, cas_row);
@@ -427,11 +435,9 @@
 			dwt_decode_1(bj, dn, sn, cas_row);
 			for (k = 0; k < rw; k++)  aj[k] = bj[k];
 		}
-		opj_free(bj);
 		
 		sn = rh1;
 		dn = rh - rh1;
-		bj = (int*)opj_malloc(rh * sizeof(int));
 		for (j = 0; j < rw; j++) {
 			aj = a + j;
 			dwt_interleave_v(aj, bj, dn, sn, w, cas_col);
@@ -438,8 +444,8 @@
 			dwt_decode_1(bj, dn, sn, cas_col);
 			for (k = 0; k < rh; k++)  aj[k * w] = bj[k];
 		}
-		opj_free(bj);
 	}
+	dwt_decode_free(bj);
 }
 
 
@@ -531,6 +537,7 @@
 	w = tilec->x1-tilec->x0;
 	l = tilec->numresolutions-1;
 	a = tilec->data;
+	bj = dwt_decode_alloc(tilec, stop);
 	
 	for (i = l-1; i >= stop; i--) {
 		int rw;			/* width of the resolution level computed                       */
@@ -551,7 +558,6 @@
         
 		sn = rw1;
 		dn = rw-rw1;
-		bj = (int*)opj_malloc(rw * sizeof(int));
 		for (j = 0; j < rh; j++) {
 			aj = a + j * w;
 			dwt_interleave_h(aj, bj, dn, sn, cas_col);
@@ -558,11 +564,9 @@
 			dwt_decode_1_real(bj, dn, sn, cas_col);
 			for (k = 0; k < rw; k++)  aj[k] = bj[k];
 		}
-		opj_free(bj);
 		
 		sn = rh1;
 		dn = rh-rh1;
-		bj = (int*)opj_malloc(rh * sizeof(int));
 		for (j = 0; j < rw; j++) {
 			aj = a + j;
 			dwt_interleave_v(aj, bj, dn, sn, w, cas_row);
@@ -569,8 +573,8 @@
 			dwt_decode_1_real(bj, dn, sn, cas_row);
 			for (k = 0; k < rh; k++)  aj[k * w] = bj[k];
 		}
-		opj_free(bj);
 	}
+	dwt_decode_free(bj);
 }
 
 
@@ -609,3 +613,30 @@
 		dwt_encode_stepsize((int) floor(stepsize * 8192.0), prec + gain, &tccp->stepsizes[bandno]);
 	}
 }
+
+
+/* <summary>                             */
+/* Allocate memory for maximum computed resolution level in decode */
+/* </summary>                            */
+static int* dwt_decode_alloc(opj_tcd_tilecomp_t * tilec, int stop) {
+	opj_tcd_resolution_t* tr = tilec->resolutions;
+	int i	= tilec->numresolutions - stop;
+	int mr	= 1;
+	int w;
+	while( --i ) {
+		tr++;
+		if( mr < ( w = tr->x1 - tr->x0 ) )
+			mr = w ;
+		if( mr < ( w = tr->y1 - tr->y0 ) )
+			mr = w ;
+	}
+	return (int*)opj_malloc(mr * sizeof(int));
+}
+
+
+/* <summary>                             */
+/* Free memory used for maximum computed resolution level in decode */
+/* </summary>                            */
+static void dwt_decode_free(int* m) {
+	opj_free(m);
+}