ref: 8a5ad7cc54f23fb99e231be3a123d6488829dc9e
parent: 190fc319c5a87c1a971ba0efa8238d9e1c88ea1f
author: Robin Watts <Robin.Watts@artifex.com>
date: Thu Jan 23 07:09:08 EST 2020
jbig2dec: Optimise jbig2_decode_generic_template0_TPGDON Decodes of JBig2_042_08.pdf (the worst comparing file to luratech in terms of time) show significant time in that function (45%). With this patch that drops to 29%.
--- a/jbig2_generic.c
+++ b/jbig2_generic.c
@@ -831,6 +831,8 @@
bool bit;
int LTP = 0;
int code = 0;
+ int gmin, gmax;
+ uint32_t left, right, top;
if (pixel_outside_field(params->gbat[0], params->gbat[1]) ||
pixel_outside_field(params->gbat[2], params->gbat[3]) ||
@@ -839,6 +841,45 @@
return jbig2_error(ctx, JBIG2_SEVERITY_FATAL, segment->number,
"adaptive template pixel is out of field");
+ /* We divide the width into 3 regions 0..left...right...GBW,
+ * between left and right, we know that our accesses will never
+ * step outside the image, enabling us to use faster accessors. */
+ left = 4;
+ right = 2;
+ gmin = gmax = params->gbat[0];
+ if (params->gbat[2] < gmin)
+ gmin = params->gbat[2];
+ if (gmax < params->gbat[2])
+ gmax = params->gbat[2];
+ if (params->gbat[4] < gmin)
+ gmin = params->gbat[4];
+ if (gmax < params->gbat[4])
+ gmax = params->gbat[4];
+ if (params->gbat[6] < gmin)
+ gmin = params->gbat[6];
+ if (gmax < params->gbat[6])
+ gmax = params->gbat[6];
+ if (left < -gmin)
+ left = -gmin;
+ if (right < gmax)
+ right = gmax;
+ right = GBW - right;
+ /* So 0 <= x < left or right <= x < GBW needs bounds checking. */
+
+ /* Now we do the same for the height, but here there is no bottom
+ * region, as we only ever look up for y. */
+ top = 2;
+ gmin = params->gbat[1];
+ if (params->gbat[3] < gmin)
+ gmin = params->gbat[3];
+ if (params->gbat[5] < gmin)
+ gmin = params->gbat[5];
+ if (params->gbat[7] < gmin)
+ gmin = params->gbat[7];
+ if (top < -gmin)
+ top = -gmin;
+ /* So 0 <= y < top needs bounds checking. */
+
for (y = 0; y < GBH; y++) {LTP ^= jbig2_arith_decode(as, &GB_stats[0x9B25], &code);
if (code)
@@ -846,29 +887,57 @@
if (!LTP) { for (x = 0; x < GBW; x++) { if (params->USESKIP && jbig2_image_get_pixel(params->SKIP, x, y)) {- jbig2_image_set_pixel(image, x, y, 0);
+ jbig2_image_set_pixel_fast(image, x, y, 0);
continue;
}
- CONTEXT = jbig2_image_get_pixel(image, x - 1, y);
- CONTEXT |= jbig2_image_get_pixel(image, x - 2, y) << 1;
- CONTEXT |= jbig2_image_get_pixel(image, x - 3, y) << 2;
- CONTEXT |= jbig2_image_get_pixel(image, x - 4, y) << 3;
- CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[0], y + params->gbat[1]) << 4;
- CONTEXT |= jbig2_image_get_pixel(image, x + 2, y - 1) << 5;
- CONTEXT |= jbig2_image_get_pixel(image, x + 1, y - 1) << 6;
- CONTEXT |= jbig2_image_get_pixel(image, x, y - 1) << 7;
- CONTEXT |= jbig2_image_get_pixel(image, x - 1, y - 1) << 8;
- CONTEXT |= jbig2_image_get_pixel(image, x - 2, y - 1) << 9;
- CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[2], y + params->gbat[3]) << 10;
- CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[4], y + params->gbat[5]) << 11;
- CONTEXT |= jbig2_image_get_pixel(image, x + 1, y - 2) << 12;
- CONTEXT |= jbig2_image_get_pixel(image, x, y - 2) << 13;
- CONTEXT |= jbig2_image_get_pixel(image, x - 1, y - 2) << 14;
- CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[6], y + params->gbat[7]) << 15;
+ if (y >= top && x >= left && x < right)
+ {+ CONTEXT = jbig2_image_get_pixel_fast(image, x - 1, y);
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x - 2, y) << 1;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x - 3, y) << 2;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x - 4, y) << 3;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + params->gbat[0], y + params->gbat[1]) << 4;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + 2, y - 1) << 5;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + 1, y - 1) << 6;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x, y - 1) << 7;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x - 1, y - 1) << 8;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x - 2, y - 1) << 9;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + params->gbat[2], y + params->gbat[3]) << 10;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + params->gbat[4], y + params->gbat[5]) << 11;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + 1, y - 2) << 12;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x, y - 2) << 13;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x - 1, y - 2) << 14;
+ CONTEXT |= jbig2_image_get_pixel_fast(image, x + params->gbat[6], y + params->gbat[7]) << 15;
+ }
+ else
+ {+ CONTEXT = jbig2_image_get_pixel(image, x - 1, y);
+ CONTEXT |= jbig2_image_get_pixel(image, x - 2, y) << 1;
+ CONTEXT |= jbig2_image_get_pixel(image, x - 3, y) << 2;
+ CONTEXT |= jbig2_image_get_pixel(image, x - 4, y) << 3;
+ CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[0], y + params->gbat[1]) << 4;
+ if (y >= 1)
+ {+ CONTEXT |= jbig2_image_get_pixel(image, x + 2, y - 1) << 5;
+ CONTEXT |= jbig2_image_get_pixel(image, x + 1, y - 1) << 6;
+ CONTEXT |= jbig2_image_get_pixel(image, x, y - 1) << 7;
+ CONTEXT |= jbig2_image_get_pixel(image, x - 1, y - 1) << 8;
+ CONTEXT |= jbig2_image_get_pixel(image, x - 2, y - 1) << 9;
+ }
+ CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[2], y + params->gbat[3]) << 10;
+ CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[4], y + params->gbat[5]) << 11;
+ if (y >= 2)
+ {+ CONTEXT |= jbig2_image_get_pixel(image, x + 1, y - 2) << 12;
+ CONTEXT |= jbig2_image_get_pixel(image, x, y - 2) << 13;
+ CONTEXT |= jbig2_image_get_pixel(image, x - 1, y - 2) << 14;
+ }
+ CONTEXT |= jbig2_image_get_pixel(image, x + params->gbat[6], y + params->gbat[7]) << 15;
+ }
bit = jbig2_arith_decode(as, &GB_stats[CONTEXT], &code);
if (code)
return jbig2_error(ctx, JBIG2_SEVERITY_FATAL, segment->number, "failed to decode arithmetic code when handling generic template0 TPGDON2");
- jbig2_image_set_pixel(image, x, y, bit);
+ jbig2_image_set_pixel_fast(image, x, y, bit);
}
} else {copy_prev_row(image, y);
--- a/jbig2_image.h
+++ b/jbig2_image.h
@@ -39,4 +39,28 @@
int jbig2_image_get_pixel(Jbig2Image *image, int x, int y);
void jbig2_image_set_pixel(Jbig2Image *image, int x, int y, bool value);
+static inline int
+jbig2_image_get_pixel_fast(Jbig2Image *image, int x, int y)
+{+ const int byte = (x >> 3) + y * image->stride;
+ const int bit = 7 - (x & 7);
+
+ return ((image->data[byte] >> bit) & 1);
+}
+
+/* set an individual pixel value in an image */
+static inline void
+jbig2_image_set_pixel_fast(Jbig2Image *image, int x, int y, bool value)
+{+ int scratch, mask;
+ int bit, byte;
+
+ byte = (x >> 3) + y * image->stride;
+ bit = 7 - (x & 7);
+ mask = (1 << bit) ^ 0xff;
+
+ scratch = image->data[byte] & mask;
+ image->data[byte] = scratch | (value << bit);
+}
+
#endif /* _JBIG2_IMAGE_H */