Duran
Posts: 37
|
Re:Slicing text corrumtion with SSE optimizations 6 Years ago
|
|
Stefan,
None of these chnages has fixed the problem.
Joaquim Duran
|
|
|
|
|
The administrator has disabled public write access.
|
Stefan
Posts: 91
|
Re:Slicing text corrumtion with SSE optimizations 6 Years ago
|
|
Hi Joaquim,
hmm it seems that our assembler may be the problem. Could You try to erase all '-D__HAVE_SSE__' from the SConstruct? This will add -msseX and -mfpmath=sse without triggering the assembler routines. The problem is, that we don't have a Phenom at hand and are not able to reproduce the problem.
kind regards
Stefan
|
|
|
|
|
The administrator has disabled public write access.
|
Duran
Posts: 37
|
Re:Slicing text corrumtion with SSE optimizations 6 Years ago
|
|
Stefan,
This test has been sucessfull: the slice text has not been corrupted.
If I can help to you in some way (running test programs,....), let me know.
Joaquim Duran
|
|
|
|
|
The administrator has disabled public write access.
|
Stefan
Posts: 91
|
Re:Slicing text corrumtion with SSE optimizations 6 Years ago
|
|
Hi Joaquim,
Now lets try to narrow down the responsible routine. I am pretty confident, that the culprit is mmssrc/mmsgui/fb/fillrectangle/fb_blit_blend_argb_to_yv12.cpp. Could you bring your SConscript back to normal, an change the #ifdef __HAVE_SSE__ lines in the mentioned cpp to #if 0? If its still running smoothly we have the exact routine.
kind regards
Stefan
|
|
|
|
|
The administrator has disabled public write access.
|
Duran
Posts: 37
|
Re:Slicing text corrumtion with SSE optimizations 5 Years, 12 Months ago
|
|
Hi Stefan,
I have'nt found the file with the path that you specify. I've llok for similar files that I've replaced #ifdef __HAVE_SSE__ lines in the mentioned cpp to #if 0:
src/mmsgui/fb/blit/mmsfb_blit_blend_argb_to_yv12.cpp NOK
src/mmsgui/fb/fillrectangle/mmsfb_fillrectangle_argb.cpp NOK, but better
src/mmsgui/fb/fillrectangle/mmsfb_fillrectangle_blend_argb.cpp NOK
Also, with no modifications to library, if compiled with profile, then it runs smothly, but without profiling, the slicing fails.
NOTE: I've never explained in which consist the corruption of slicing text: when slicing the text, the current text is not removed but redraw in the next position. So, alter several secons, you get a white (this is the color of the text) rectangle.
NOTE2: in the profiling, I haven't get profile information about the library, but test program. Could you check if it is passed the option '-pg' to the linker?
Joaquim Duran
|
|
|
|
|
The administrator has disabled public write access.
|
Duran
Posts: 37
|
Re:Slicing text corrumtion with SSE optimizations 5 Years, 12 Months ago
|
|
Hello,
Based in the behaviour of the profiler, I've decided to check what happens when the library is compiled using a different level of optimization. Here are the results:
-O0: OK
-O1: NOK Segmentation fault
Program received signal SIGSEGV, Segmentation fault.
0x00007ffff7aacbb5 in mmsfb_fillrectangle_rgb32 (
dst_planes=<value optimized out>, dst_height=<value optimized out>, dx=0,
dy=0, dw=800, dh=600, color=
{r = 0 '\0', g = 0 '\0', b = 0 '\0', a = 0 '\0'})
at src/mmsgui/fb/fillrectangle/mmsfb_fillrectangle_rgb32.cpp:67
67 __asm__ __volatile__ ( "\trep stosl\n" : : "D" (dst), "a" (SRC), "c" (dw));
-O2: NOK: slicing text corrupted
-O3: NOK: slicing text corrupted
The compiler is not working fine.
Joaquim Duran
|
|
|
|
|
The administrator has disabled public write access.
|
|