Active Topics
-
Maemo repository (4)
to Community by teroyk - 1 day, 4 hrs ago -
Error after installing Leste on sdcard (13)
to Maemo 7 / Leste by Maemish - 2 days, 10 hrs ago -
TLS 1.2 for N9 is ready (29)
to MeeGo / Harmattan by smatkovi - 2 days, 19 hrs ago -
Camera phone competition March 2023 "Endless" (35)
to General by Maemish - 6 days, 4 hrs ago - more...
|
2010-04-23
, 09:05
|
|
Posts: 62 |
Thanked: 97 times |
Joined on Dec 2009
@ Finland, Kerava
|
#2
|
Where could I get the code, so that I can investigate the problem?
|
|
2010-04-23
, 17:22
|
|
Posts: 46 |
Thanked: 392 times |
Joined on Apr 2010
@ Stanford University
|
#3
|
The following sample code compiles fine under scratchbox with
It doesn't compile under MADDE with
test.cpp:
(Note for those interested in NEON assembly: The above is not fast at all, so don't use it as an example! It's just a quick testcase for compilation)
Code:
g++ -mfpu=neon -c test.cpp
Code:
mad g++ -mfpu=neon -c test.cpp
Code:
#include <math.h>
#include <string.h>
#include <iostream>
#include <map>
#include <vector>
using namespace std;
void correctColors_asm(int &red, int &grn, int &blu, int* col_iColorMatrix)
{
int tmp1, tmp2, tmp3;
//Load color values into Q0-Q3 [which are equivalent to D0 - D7]
__asm__ volatile("VMOV D0, %[rval], %[rval]\n\t"
"VMOV D1, %[rval], %[rval]\n\t"
"VMOV D2, %[gval], %[gval]\n\t"
"VMOV D3, %[gval], %[gval]\n\t"
"VMOV D4, %[bval], %[bval]\n\t"
"VMOV D5, %[bval], %[bval]\n\t"
"VMOV D6, %[v1], %[v1]\n\t"
"VMOV D7, %[v1], %[v1]\n\t" :: [rval] "r" (red), [gval] "r" (grn), [bval] "r" (blu), [v1] "r" (1) : "d0", "d1", "d2", "d3", "d4", "d5", "d6", "d7");
//Load color correction matrix into Q4-Q7 [which are equivalent to D8 - D15]
__asm__ volatile("MOV r0, %[matrixPtr]" : : [matrixPtr] "r" (col_iColorMatrix) : "r0");
__asm__ volatile("VLD1.64 {D8}, [r0]!" ::: "r0", "d8");
__asm__ volatile("VLD1.32 {D9[0]}, [r0]!" ::: "r0", "d9");
__asm__ volatile("VLD1.64 {D10}, [r0]!" ::: "r0", "d10");
__asm__ volatile("VLD1.32 {D11[0]}, [r0]!" ::: "r0", "d11");
__asm__ volatile("VLD1.64 {D12}, [r0]!" ::: "r0", "d12");
__asm__ volatile("VLD1.32 {D13[0]}, [r0]!" ::: "r0", "d13");
__asm__ volatile("VLD1.64 {D14}, [r0]!" ::: "r0", "d14");
__asm__ volatile("VLD1.32 {D15[0]}, [r0]!" ::: "r0", "d15");
//'matrix' mutliply the vectors together
__asm__ volatile("VMUL.I32 Q3, Q3, Q7\n\t"
"VMLA.I32 Q3, Q0, Q4\n\t"
"VMLA.I32 Q3, Q1, Q5\n\t"
"VMLA.I32 Q3, Q2, Q6" ::: "q3");
//ARM has a mov with a right shift built in as one instruction, so right shifting the vector is a waste of time
//__asm__ volatile("VSHR.S32 Q3, Q3, #10" ::: "q3");
//extract final rgb values from vector
__asm__ volatile("VMOV %[tmp1], %[tmp2], D6" : [tmp1] "=r" (tmp1), [tmp2] "=r" (tmp2) :);
__asm__ volatile("VMOV %[tmp3], r0, D7" : [tmp3] "=r" (tmp3) :: "r0");
red = tmp1;
grn = tmp2;
blu = tmp3;
}
|
|
2010-04-25
, 23:02
|
|
Posts: 1 |
Thanked: 0 times |
Joined on Apr 2010
|
#4
|
|
|
2010-04-26
, 07:20
|
|
Posts: 62 |
Thanked: 97 times |
Joined on Dec 2009
@ Finland, Kerava
|
#5
|
Originally Posted by jlebar
Yes, this bug is relevant.
Perhaps this bug is relevant?
https://bugs.maemo.org/show_bug.cgi?id=9624
|
|
2010-04-27
, 19:13
|
|
Posts: 46 |
Thanked: 392 times |
Joined on Apr 2010
@ Stanford University
|
#6
|
Great, thanks. That was the issue, and the fix listed in the bug took care of it.
The library is a very simple Makefile-based system, so I tried to build it in MADDE with a 'mad make'. This mostly worked fine, but then it ran into a file with some inline NEON assembly. The result is:
....
g++ -O3 -Wall -mfpu=neon -mfloat-abi=softfp -c processing/Demosaic.cpp -o processing/Demosaic.o C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s: Assembler messages: C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:51: Error: bad instruction `vld1.64 {D8},[r0]!' C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:52: Error: bad instruction `vld1.32 {D9[0]},[r0]!' C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:53: Error: bad instruction `vld1.64 {D10},[r0]!' C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:54: Error: bad instruction `vld1.32 {D11[0]},[r0]!' C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:55: Error: bad instruction `vld1.64 {D12},[r0]!' C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:56: Error: bad instruction `vld1.32 {D13[0]},[r0]!' C:/Users/talvala/AppData/Local/Temp/ccYQ7YcJ.s:57: Error: bad instruction `vld1.64 {D14},[r0]!'and so on.
The g++ version in MADDE ('mad g++ --version') seems to match the scratchbox version (4.2.1), so I'm at a bit of a loss as to why this won't work.
Anyone have any ideas?