Speed up shadow generation for small windows

Signed-off-by: Yuxuan Shui <yshuiv7@gmail.com>

src/backend/backend_common.c

@@ -47,7 +47,7 @@ solid_picture(session_t *ps, bool argb, double a, double r, double g, double b)
 }
 
 xcb_image_t *make_shadow(xcb_connection_t *c, const conv *kernel,
-                         const double *shadow_sum, double opacity, int width, int height) {
+                         double opacity, int width, int height) {
 	/*
 	 * We classify shadows into 4 kinds of regions
 	 *    r = shadow radius
@@ -62,6 +62,7 @@ xcb_image_t *make_shadow(xcb_connection_t *c, const conv *kernel,
 	 * height+r +-----+---------+-----+
 	 */
 	xcb_image_t *ximage;
+	const double *shadow_sum = kernel->rsum;
 	int d = kernel->size, r = d / 2;
 	int swidth = width + r * 2, sheight = height + r * 2;
 
@@ -181,7 +182,7 @@ bool build_shadow(session_t *ps, double opacity, const int width, const int heig
 	xcb_gcontext_t gc = None;
 
 	shadow_image =
-	    make_shadow(ps->c, ps->gaussian_map, ps->shadow_sum, opacity, width, height);
+	    make_shadow(ps->c, ps->gaussian_map, opacity, width, height);
 	if (!shadow_image) {
 		log_error("Failed to make shadow");
 		return false;

src/backend/backend_common.h

@@ -19,5 +19,5 @@ solid_picture(session_t *ps, bool argb, double a, double r, double g, double b);
 
 void paint_all_new(session_t *ps, region_t *region, win *const t);
 
-xcb_image_t *make_shadow(xcb_connection_t *c, const conv *kernel,
+xcb_image_t *
-                         const double *shadow_sum, double opacity, int width, int height);
+make_shadow(xcb_connection_t *c, const conv *kernel, double opacity, int width, int height);

src/common.h

@@ -531,8 +531,6 @@ typedef struct session {
   /// Gaussian map of shadow.
   conv *gaussian_map;
   // for shadow precomputation
-  /// Pre-computed table for shadow.
-  double *shadow_sum;
   /// A region in which shadow is not painted on.
   region_t shadow_exclude_reg;
 

src/compton.c

@@ -2640,7 +2640,6 @@ session_init(session_t *ps_old, int argc, char **argv) {
     .cshadow_picture = XCB_NONE,
     .white_picture = XCB_NONE,
     .gaussian_map = NULL,
-    .shadow_sum = NULL,
 
     .refresh_rate = 0,
     .refresh_intv = 0UL,

src/kernel.c

@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: MPL-2.0
 // Copyright (c) Yuxuan Shui <yshuiv7@gmail.com>
 
+#include <assert.h>
 #include <math.h>
 
 #include "kernel.h"
@@ -28,9 +29,20 @@ double sum_kernel(const conv *map, int x, int y, int width, int height) {
 	if (yend > map->size)
 		yend = map->size;
 
+	assert(yend > 0 && xend > 0);
+
+	int d = map->size;
+	if (map->rsum) {
+		double v1 = xstart ? map->rsum[(yend - 1) * d + xstart - 1] : 0;
+		double v2 = ystart ? map->rsum[(ystart - 1) * d + xend - 1] : 0;
+		double v3 =
+		    (xstart && ystart) ? map->rsum[(ystart - 1) * d + xstart - 1] : 0;
+		return map->rsum[(yend - 1) * d + xend - 1] - v1 - v2 + v3;
+	}
+
 	for (int yi = ystart; yi < yend; yi++) {
 		for (int xi = xstart; xi < xend; xi++) {
-			ret += map->data[yi * map->size + xi];
+			ret += map->data[yi * d + xi];
 		}
 	}
 
@@ -65,6 +77,7 @@ conv *gaussian_kernel(double r) {
 
 	c = cvalloc(sizeof(conv) + size * size * sizeof(double));
 	c->size = size;
+	c->rsum = NULL;
 	t = 0.0;
 
 	for (int y = 0; y < size; y++) {
@@ -86,13 +99,13 @@ conv *gaussian_kernel(double r) {
 
 /// preprocess kernels to make shadow generation faster
 /// shadow_sum[x*d+y] is the sum of the kernel from (0, 0) to (x, y), inclusive
-void shadow_preprocess(conv *map, double **shadow_sum) {
+void shadow_preprocess(conv *map) {
 	const int d = map->size;
 
-	if (*shadow_sum)
+	if (map->rsum)
-		free(*shadow_sum);
+		free(map->rsum);
 
-	auto sum = *shadow_sum = ccalloc(d * d, double);
+	auto sum = map->rsum = ccalloc(d * d, double);
 	sum[0] = map->data[0];
 
 	for (int x = 1; x < d; x++) {

src/kernel.h

@@ -2,12 +2,14 @@
 // Copyright (c) Yuxuan Shui <yshuiv7@gmail.com>
 
 #pragma once
+#include <stdlib.h>
 #include "compiler.h"
 
 /// Code for generating convolution kernels
 
 typedef struct conv {
 	int size;
+	double *rsum;
 	double data[];
 } conv;
 
@@ -21,4 +23,9 @@ conv *gaussian_kernel(double r);
 
 /// preprocess kernels to make shadow generation faster
 /// shadow_sum[x*d+y] is the sum of the kernel from (0, 0) to (x, y), inclusive
-void shadow_preprocess(conv *map, double **shadow_sum);
+void shadow_preprocess(conv *map);
+
+static inline void free_conv(conv *k) {
+	free(k->rsum);
+	free(k);
+}

src/render.c

@@ -451,7 +451,7 @@ static bool win_build_shadow(session_t *ps, win *w, double opacity) {
 	xcb_gcontext_t gc = XCB_NONE;
 
 	shadow_image =
-	    make_shadow(ps->c, ps->gaussian_map, ps->shadow_sum, opacity, width, height);
+	    make_shadow(ps->c, ps->gaussian_map, opacity, width, height);
 	if (!shadow_image) {
 		log_error("failed to make shadow");
 		return XCB_NONE;
@@ -1058,7 +1058,7 @@ bool init_render(session_t *ps) {
 	}
 
 	ps->gaussian_map = gaussian_kernel(ps->o.shadow_radius);
-	shadow_preprocess(ps->gaussian_map, &ps->shadow_sum);
+	shadow_preprocess(ps->gaussian_map);
 
 	ps->black_picture = solid_picture(ps, true, 1, 0, 0, 0);
 	ps->white_picture = solid_picture(ps, true, 1, 1, 1, 1);
@@ -1116,8 +1116,7 @@ void deinit_render(session_t *ps) {
 
 	free_picture(ps->c, &ps->black_picture);
 	free_picture(ps->c, &ps->white_picture);
-	free(ps->shadow_sum);
+	free_conv(ps->gaussian_map);
-	free(ps->gaussian_map);
 
 	// Free other X resources
 	free_root_tile(ps);