summaryrefslogtreecommitdiffstats
path: root/meta/packages/xserver/xserver-kdrive/faster-rotated.patch
diff options
context:
space:
mode:
Diffstat (limited to 'meta/packages/xserver/xserver-kdrive/faster-rotated.patch')
-rw-r--r--meta/packages/xserver/xserver-kdrive/faster-rotated.patch241
1 files changed, 241 insertions, 0 deletions
diff --git a/meta/packages/xserver/xserver-kdrive/faster-rotated.patch b/meta/packages/xserver/xserver-kdrive/faster-rotated.patch
new file mode 100644
index 0000000000..eaf7ddec36
--- /dev/null
+++ b/meta/packages/xserver/xserver-kdrive/faster-rotated.patch
@@ -0,0 +1,241 @@
1Index: xserver/miext/shadow/shrotate.c
2===================================================================
3RCS file: /scratch/openbsd/cvs/XF4/xc/programs/Xserver/miext/shadow/shrotate.c,v
4retrieving revision 1.2
5diff -u -r1.2 shrotate.c
6--- xserver/miext/shadow/shrotate.c 3 Nov 2004 00:09:54 -0000 1.2
7+++ xserver/miext/shadow/shrotate.c 20 Sep 2005 23:07:58 -0000
8@@ -45,6 +45,106 @@
9 #define TOP_TO_BOTTOM 2
10 #define BOTTOM_TO_TOP -2
11
12+
13+static void
14+shadowUpdateRotatePackedSubRectangle(shadowBufPtr pBuf,
15+ FbBits *shaLine, int shaFirstShift,
16+ int shaStepOverX, int shaStepOverY,
17+ int shaStepDownX, int shaStepDownY,
18+ int shaBpp, FbBits shaMask,
19+ ScreenPtr pScreen,
20+ int scr_x1, int scr_y,
21+ int scr_h, int scr_w,
22+ int pixelsPerBits)
23+{
24+ FbBits *sha;
25+ int shaShift;
26+ int scr_x;
27+ int w;
28+
29+ /*
30+ * Copy the bits, always write across the physical frame buffer
31+ * to take advantage of write combining.
32+ */
33+ while (scr_h--)
34+ {
35+ int p;
36+ FbBits bits;
37+ FbBits *win;
38+ int i;
39+ CARD32 winSize;
40+
41+ sha = shaLine;
42+ shaShift = shaFirstShift;
43+ w = scr_w;
44+ scr_x = scr_x1 * shaBpp >> FB_SHIFT;
45+
46+ while (w)
47+ {
48+ /*
49+ * Map some of this line
50+ */
51+ win = (FbBits *) (*pBuf->window) (pScreen,
52+ scr_y,
53+ scr_x << 2,
54+ SHADOW_WINDOW_WRITE,
55+ &winSize,
56+ pBuf->closure);
57+ i = (winSize >> 2);
58+ if (i > w)
59+ i = w;
60+ w -= i;
61+ scr_x += i;
62+ /*
63+ * Copy the portion of the line mapped
64+ */
65+ while (i--)
66+ {
67+ bits = 0;
68+ p = pixelsPerBits;
69+ /*
70+ * Build one word of output from multiple inputs
71+ */
72+ while (p--)
73+ {
74+ bits = FbScrLeft(bits, shaBpp);
75+ bits |= FbScrRight (*sha, shaShift) & shaMask;
76+
77+ shaShift -= shaStepOverX;
78+ if (shaShift >= FB_UNIT)
79+ {
80+ shaShift -= FB_UNIT;
81+ sha--;
82+ }
83+ else if (shaShift < 0)
84+ {
85+ shaShift += FB_UNIT;
86+ sha++;
87+ }
88+ sha += shaStepOverY;
89+ }
90+ *win++ = bits;
91+ }
92+ }
93+ scr_y++;
94+ shaFirstShift -= shaStepDownX;
95+ if (shaFirstShift >= FB_UNIT)
96+ {
97+ shaFirstShift -= FB_UNIT;
98+ shaLine--;
99+ }
100+ else if (shaFirstShift < 0)
101+ {
102+ shaFirstShift += FB_UNIT;
103+ shaLine++;
104+ }
105+ shaLine += shaStepDownY;
106+ }
107+}
108+
109+#define BLOCKSIZE_HEIGHT 32
110+#define BLOCKSIZE_WIDTH 32
111+
112 void
113 shadowUpdateRotatePacked (ScreenPtr pScreen,
114 shadowBufPtr pBuf)
115@@ -61,7 +161,6 @@
116 int sha_x1 = 0, sha_y1 = 0;
117 int scr_x1 = 0, scr_x2 = 0, scr_y1 = 0, scr_y2 = 0, scr_w, scr_h;
118 int scr_x, scr_y;
119- int w;
120 int pixelsPerBits;
121 int pixelsMask;
122 FbStride shaStepOverY = 0, shaStepDownY = 0;
123@@ -221,86 +320,46 @@
124 ((sha_x1 * shaBpp) >> FB_SHIFT));
125
126 /*
127- * Copy the bits, always write across the physical frame buffer
128- * to take advantage of write combining.
129+ * Copy in blocks of size BLOCKSIZE_WIDTH x BLOCKSIZE_HEIGHT
130+ * to reduce the number of cache misses when rotating 90 or
131+ * 270 degrees.
132 */
133- while (scr_h--)
134+ for (scr_y = scr_y1; scr_y < scr_y2; scr_y += BLOCKSIZE_HEIGHT)
135 {
136- int p;
137- FbBits bits;
138- FbBits *win;
139- int i;
140- CARD32 winSize;
141-
142 sha = shaLine;
143 shaShift = shaFirstShift;
144- w = scr_w;
145- scr_x = scr_x1 * shaBpp >> FB_SHIFT;
146
147- while (w)
148+ for (scr_x = scr_x1; scr_x < scr_x2; scr_x += BLOCKSIZE_WIDTH)
149 {
150- /*
151- * Map some of this line
152- */
153- win = (FbBits *) (*pBuf->window) (pScreen,
154- scr_y,
155- scr_x << 2,
156- SHADOW_WINDOW_WRITE,
157- &winSize,
158- pBuf->closure);
159- i = (winSize >> 2);
160- if (i > w)
161- i = w;
162- w -= i;
163- scr_x += i;
164- /*
165- * Copy the portion of the line mapped
166- */
167- while (i--)
168- {
169- bits = 0;
170- p = pixelsPerBits;
171- /*
172- * Build one word of output from multiple inputs
173- *
174- * Note that for 90/270 rotations, this will walk
175- * down the shadow hitting each scanline once.
176- * This is probably not very efficient.
177- */
178- while (p--)
179- {
180- bits = FbScrLeft(bits, shaBpp);
181- bits |= FbScrRight (*sha, shaShift) & shaMask;
182+ int h = BLOCKSIZE_HEIGHT;
183+ int w = BLOCKSIZE_WIDTH;
184
185- shaShift -= shaStepOverX;
186- if (shaShift >= FB_UNIT)
187- {
188- shaShift -= FB_UNIT;
189- sha--;
190- }
191- else if (shaShift < 0)
192- {
193- shaShift += FB_UNIT;
194- sha++;
195- }
196- sha += shaStepOverY;
197- }
198- *win++ = bits;
199- }
200- }
201- scr_y++;
202- shaFirstShift -= shaStepDownX;
203- if (shaFirstShift >= FB_UNIT)
204- {
205- shaFirstShift -= FB_UNIT;
206- shaLine--;
207- }
208- else if (shaFirstShift < 0)
209- {
210- shaFirstShift += FB_UNIT;
211- shaLine++;
212+ if (scr_y + h > scr_y2)
213+ h = scr_y2 - scr_y;
214+ if (scr_x + w > scr_x2)
215+ w = scr_x2 - scr_x;
216+ w = (w * shaBpp) >> FB_SHIFT;
217+
218+ shadowUpdateRotatePackedSubRectangle
219+ (pBuf,
220+ sha, shaShift,
221+ shaStepOverX, shaStepOverY,
222+ shaStepDownX, shaStepDownY,
223+ shaBpp, shaMask,
224+ pScreen,
225+ scr_x, scr_y,
226+ h, w,
227+ pixelsPerBits);
228+
229+ shaShift -= BLOCKSIZE_WIDTH * shaStepOverX;
230+ sha += BLOCKSIZE_WIDTH * shaStepOverY;
231+ sha -= (shaShift >> FB_SHIFT);
232+ shaShift &= FB_MASK;
233 }
234- shaLine += shaStepDownY;
235+ shaFirstShift -= BLOCKSIZE_HEIGHT * shaStepDownX;
236+ shaLine += BLOCKSIZE_HEIGHT * shaStepDownY;
237+ shaLine -= (shaFirstShift >> FB_SHIFT);
238+ shaFirstShift &= FB_MASK;
239 }
240 }
241 }
generated by cgit v1.2.3-54-g00ecf (git 2.39.0) at 2024-05-02 00:42:20 +0000