jpeg2000: Optimize dequantization

Float:   4700 -> 2700 cycles
Integer: 4400 -> 2800 cycles

(sandybridge  i7)
Signed-off-by: Luca Barbato <lu_zero@gentoo.org>

libavcodec/jpeg2000dec.c

@@ -1006,13 +1006,14 @@ static void dequantization_float(int x, int y, Jpeg2000Cblk *cblk,
                                  Jpeg2000Component *comp,
                                  Jpeg2000T1Context *t1, Jpeg2000Band *band)
 {
-    int i, j, idx;
-    float *datap = &comp->f_data[(comp->coord[0][1] - comp->coord[0][0]) * y + x];
-    for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j)
-        for (i = 0; i < (cblk->coord[0][1] - cblk->coord[0][0]); ++i) {
-            idx        = (comp->coord[0][1] - comp->coord[0][0]) * j + i;
-            datap[idx] = (float)(t1->data[j][i]) * band->f_stepsize;
-        }
+    int i, j;
+    int w = cblk->coord[0][1] - cblk->coord[0][0];
+    for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j) {
+        float *datap = &comp->f_data[(comp->coord[0][1] - comp->coord[0][0]) * (y + j) + x];
+        int *src = t1->data[j];
+        for (i = 0; i < w; ++i)
+            datap[i] = src[i] * band->f_stepsize;
+    }
 }
 
 /* Integer dequantization of a codeblock.*/
@@ -1020,14 +1021,14 @@ static void dequantization_int(int x, int y, Jpeg2000Cblk *cblk,
                                Jpeg2000Component *comp,
                                Jpeg2000T1Context *t1, Jpeg2000Band *band)
 {
-    int i, j, idx;
-    int32_t *datap = &comp->i_data[(comp->coord[0][1] - comp->coord[0][0]) * y + x];
-    for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j)
-        for (i = 0; i < (cblk->coord[0][1] - cblk->coord[0][0]); ++i) {
-            idx        = (comp->coord[0][1] - comp->coord[0][0]) * j + i;
-            datap[idx] =
-                ((int32_t)(t1->data[j][i]) * band->i_stepsize + (1 << 15)) >> 16;
-        }
+    int i, j;
+    int w = cblk->coord[0][1] - cblk->coord[0][0];
+    for (j = 0; j < (cblk->coord[1][1] - cblk->coord[1][0]); ++j) {
+        int32_t *datap = &comp->i_data[(comp->coord[0][1] - comp->coord[0][0]) * (y + j) + x];
+        int *src = t1->data[j];
+        for (i = 0; i < w; ++i)
+            datap[i] = (src[i] * band->i_stepsize + (1 << 15)) >> 16;
+    }
 }
 
 /* Inverse ICT parameters in float and integer.