So in the Halide aot example https://halide-lang.org/tutorials/tutorial_lesson_10_aot_compilation_run.html, there's this line:
Halide::Runtime::Buffer<uint8_t> input(640, 480), output(640, 480);
My question is, how does one load an image into the input runtime buffer?
The same way the previous tutorials load images.
Add this include:
#include "halide_image_io.h" // for load_image and save_image
Then replace
Halide::Runtime::Buffer<uint8_t> input(640, 480), output(640, 480);
with
Halide::Runtime::Buffer<uint8_t> input = Halide::Tools::load_image("path/to/input.png");
Halide::Runtime::Buffer<uint8_t> output(input.width(), input.height());
If you're interested in saving the output, after the error check, add the line:
Halide::Tools::save_image(output, "path/to/output.png");
Note that the paths, if not absolute paths, will be in the following directories: the Halide/tutorial/ directory for input, and the Halide/bin/build/tmp/ directory for the output, after running:
make tutorial_lesson_10_aot_compilation_run
from the root directory of Halide.
The latest syntax is as follows:
Halide::Runtime::Buffer<float> in = Tools::load_image("image.png");
Halide::Runtime::Buffer<float> out(in.width(), in.height(),in.channels());
int error = halide_blur(in, out); //halide_blur is the function defined in aot compiled code
Tools::save_image(out, output.png");
Related
I'm trying to add a simple line in fstab within
the final rootfs that Yocto builds.
My first approach was to add my own fstab in my layer meta-mylayer/recipes-core/base-files/base-files/fstab and the proper meta-mylayer/recipes-core/base-files/base-files/base-files_%.bbappend which only have the following line:
FILESEXTRAPATHS_prepend := "${THISDIR}/${PN}:"
And it works, but as the title of my question says, i want to modify fstab based on the recipe-image i want to build i.e. dev-image & prod-image.
After some investigation i think i have 2 options
Modify fstab within the recipe image, extending the do_install task...
dev-image.bb
--------------
DESCRIPTION = "Development Image"
[...]
inherit core-image
do_install_append () {
echo "======= Modifying fstab ========"
cat >> ${D}${sysconfdir}/fstab <<EOF
# The line i want to Add
EOF
}
[...]
--------------
Problem is that i'm actually not seeing my modified line in my final /etc/fstab and bitbake is not showing any build error or warning about this, actually, i'm not even able to see the echo-trace i put.
My second attempt was to handle these modifications with packages and depending on the recipe-image i will be able to add the package for *-dev or *-prod. This idea was taken from Oleksandr Poznyak in this answer in summary he suggest the following:
1) Create *.bbappend recipe base-files_%s.bbappend in your layer. It
appends to poky "base-files" recipe.
2) Create your own "python do_package_prepend" function where you should
make your recipe produce two different packages
3) Add them to DEPENDS in your image recipe
And based on his example i made my own recipe:
base-files_%.bbappend
-------------------------
FILESEXTRAPATHS_prepend := "${THISDIR}/${PN}:"
SRC_URI += "file://fstab-dev \
file://fstab-prod \
"
PACKAGES += " ${PN}-dev ${PN}-prod"
CONFFILES_${PN}-dev = "${CONFFILES_${PN}}"
CONFFILES_${PN}-prod = "${CONFFILES_${PN}}"
pkg_preinst_${PN}-dev = "${pkg_preinst_${PN}}"
pkg_preinst_${PN}-prod = "${pkg_preinst_${PN}}"
RREPLACES_${PN}-dev = "${PN}"
RPROVIDES_${PN}-dev = "${PN}"
RCONFLICTS_${PN}-dev = "${PN}"
RREPLACES_${PN}-prod = "${PN}"
RPROVIDES_${PN}-prod = "${PN}"
RCONFLICTS_${PN}-prod = "${PN}"
python populate_packages_prepend() {
import shutil
packages = ("${PN}-dev", "${PN}-prod")
for package in packages:
# copy ${PN} content to packages
shutil.copytree("${PKGD}", "${PKGDEST}/%s" % package, symlinks=True)
# replace fstab
if package == "${PN}-dev":
shutil.copy("${WORKDIR}/fstab-dev", "${PKGDEST}/${PN}-dev/etc/fstab")
else:
shutil.copy("${WORKDIR}/fstab-prod", "${PKGDEST}/${PN}-prod/etc/fstab")
}
-------------------------
And in my recipe-image(dev-image.bb) i added base-files-dev packet
dev-image.bb
--------------
DESCRIPTION = "Development Image"
[...]
inherit core-image
IMAGE_INSTALL = " \
${MY_PACKETS} \
base-files-dev \
"
[...]
--------------
Problem with this, is that i'm not familiarized with phyton indentation so probably i'm messing things up, the error log shows as follows.
DEBUG: Executing python function populate_packages
ERROR: Error executing a python function in exec_python_func() autogenerated:
The stack trace of python calls that resulted in this exception/failure was:
File: 'exec_python_func() autogenerated', lineno: 2, function: <module>
0001:
*** 0002:populate_packages(d)
0003:
File: '/home/build/share/build_2/../sources/poky/meta/classes/package.bbclass', lineno: 1138, function: populate_packages
1134:
1135: workdir = d.getVar('WORKDIR')
1136: outdir = d.getVar('DEPLOY_DIR')
1137: dvar = d.getVar('PKGD')
*** 1138: packages = d.getVar('PACKAGES').split()
1139: pn = d.getVar('PN')
1140:
1141: bb.utils.mkdirhier(outdir)
1142: os.chdir(dvar)
File: '/usr/lib/python3.6/shutil.py', lineno: 315, function: copytree
0311: destination path as arguments. By default, copy2() is used, but any
0312: function that supports the same signature (like copy()) can be used.
0313:
0314: """
*** 0315: names = os.listdir(src)
0316: if ignore is not None:
0317: ignored_names = ignore(src, names)
0318: else:
0319: ignored_names = set()
Exception: FileNotFoundError: [Errno 2] No such file or directory: '${PKGD}'
DEBUG: Python function populate_packages finished
DEBUG: Python function do_package finished
I will really appreciate any clue or sort of direction, i'm not an Yocto expert so maybe the options that i suggest are not the most elegant and probably there is a better way to do it, so be free to give me any recommendation.
Thank you very much.
UPDATE:
As always, i was not the only one trying this, the way that i make it work was thanks this answer the only inconvenience with this is that you need to rm what you want to install through a .bbappend but for now is fine for me.
I also tried to do the same with bbclasses, which for me, it is a more elegant wayto do it, but i failed... i got the following error
ERROR: base-files-dev-3.0.14-r89 do_packagedata: The recipe base-files-dev is trying to install files into a shared area when those files already exist. Those files and their manifest location are:
I tried to rm fstab within the .bbappend but the same error is showed
Maybe somebody will share what i'm doing wrong...
If you don't find this post valuable please remove...
Your recipe which base on Oleksandr doesn't work due to dropped support for variables expansion in newer Poky.
https://www.yoctoproject.org/docs/latest/mega-manual/mega-manual.html#migration-2.1-variable-expansion-in-python-functions
Error explicit says:
Exception: FileNotFoundError: [Errno 2] No such file or directory: '${PKGD}'
It didn't expand the variable.
P.S.
This is not a proper answer to Your question but SO blocks comments.
I am trying to put together and example of coding inline assembly code in a 'C' program. I have not had any success. I am using GCC 4.9.0. As near as I can tell my syntax is correct. However the build blows up with the following errors:
/tmp/ccqC2wtq.s:48: Error: syntax error; found `(', expected `,'
Makefile:51: recipe for target 'all' failed
/tmp/ccqC2wtq.s:48: Error: junk at end of line: `(31)'
/tmp/ccqC2wtq.s:49: Error: syntax error; found `(', expected `,'
/tmp/ccqC2wtq.s:49: Error: junk at end of line: `(9)'
These are related to the input/output/clobbers lines in the code. Anyone have an idea where I went wrong?
asm volatile("li 7, %0\n" // load destination address
"li 8, %1\n" // load source address
"li 9, 100\n" // load count
// save source address for return
"mr 3,7\n"
// prepare for loop
"subi 8,8,1\n"
"subi 9,9,1\n"
// perform copy
"1:\n"
"lbzu 10,2(8)\n"
"stbu 10,2(7)\n"
"subi 9,9,1\n" // Decrement the count
"bne 1b\n" // If zero, we've finished
"blr\n"
: // no outputs
: "m" (array), "m" (stringArray)
: "r7", "r8"
);
It's not clear what you are trying to do with the initial instructions:
li 7, %0
li 8, %1
Do you want to load the address of the variables into those registers? In general, this is not possible because the address is not representable in an immediate operand. The easiest way out is to avoid using r7 and r8, and instead use %0 and %1 directly as the register names. It seems that you want to use these registers as base addresses, so you should use the b constraint:
: "b" (array), "b" (stringArray)
Then GCC will take care of the details of materializing the address in a suitable fashion.
You cannot return using blr from inline assembler because it's not possible to tear down the stack frame GCC created. You also need to double-check the clobbers and make sure that you list all the things you clobber (including condition codes, memory, and overwritten input operands).
I was able to get this working by declaring a pair of pointers and initializing them with the addresses of the arrays. I didn't realize that the addresses wouldn't be available directly. I've used inline assembly very sparsely in the past, usually just
to raise or lower interrupt masks, not to do anything that references variables. I just
had some folks who wanted an example. And the "blr" was leftover when I copied a
snipet of a pure assembly routine to use as a starting point. Thanks for the responses.
The final code piece looks like this:
int main()
{
char * stringArrayPtr;
unsigned char * myArrayPtr;
unsigned char myArray[100];
stringArrayPtr = (char *)&stringArray;
myArrayPtr = myArray;
asm volatile(
"lwz 7,%[myArrayPtr]\n" // load destination address
"lwz 8, %[stringArrayPtr]\n" // load source address
"li 9, 100\n" // load count
"mr 3,7\n" // save source address for return
"subi 8,8,1\n" // prepare for loop
"subi 9,9,1\n"
"1:\n" // perform copy
"lbzu 10,1(8)\n"
"stbu 10,1(7)\n"
"subic. 9,9,1\n" // Decrement the count
"bne 1b\n" // If zero, we've finished
// The outputs / inputs / clobbered list
// No output variables are specified here
// See GCC documentation on extended assembly for details.
:
: [stringArrayPtr] "m" (stringArrayPtr), [myArrayPtr]"m"(myArrayPtr)
: "7","8","9","10"
);
}
I cannot get what size parameters are for fftwf_plan_dft_r2c_2d
Input: a N rows by M cols matrix
Output: a N rows by floor(M/2) + 1 cols matrix ?
Are parameters input or output size?
Tried to give input size. This is what GDB sais
Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 1676.0x768]
0x637eed72 in n1fv_8 () from C:\devfiles\bin\libfftw3f-3.dll
(gdb) backtrace
#0 0x637eed72 in n1fv_8 () from C:\devfiles\bin\libfftw3f-3.dll
#1 0x7c91a000 in ntdll!RtlpUnWaitCriticalSection ()
from C:\WINDOWS\system32\ntdll.dll
No other thread uses fftw at the time.
The context:
Herbs::MatrixStorage<float> spectrum_in(frame_a.nRowsGet(),frame_a.nColsGet());
Herbs::MatrixStorage<std::complex<float>>
spectrum_out(frame_a.nRowsGet(),frame_a.nColsGet()/2+1);
Check for heap corruption issued by possible alllocation mistakes in MatrixStorage class . The rows in the matrix is one large block. rowGet returns the pointer to the given row.
memset(spectrum_in.rowGet(0),0
,sizeof(float)*spectrum_in.nRowsGet()*spectrum_in.nColsGet());
memset(spectrum_out.rowGet(0),0
,sizeof(float)*spectrum_out.nRowsGet()*spectrum_out.nColsGet());
heapdump(); //Heap seams to be fine after these
Causes sigsevg
FFT::PlanFloat_2dR2C plan(spectrum_in,spectrum_out);
The plan constructor does the following
plan=FFT::PlanFloat_2dR2C::PlanFloat_2dR2C(Herbs::MatrixStorage<InputType>& buffer_in
,Herbs::MatrixStorage<OutputType>& buffer_out)
{
plan=fftwf_plan_dft_r2c_2d
(
buffer_in.nRowsGet()
,buffer_in.nColsGet()
,buffer_in.rowGet(0)
,(fftwf_complex*)buffer_out.rowGet(0)
,FFTW_MEASURE
);
}
EDIT:
I used a precompiled DLL instead. GCC may produce bad code on 32-bit Windows (From release notes):
Removed an archaic stack-alignment hack that was failing with gcc-4.7/i386. Added stack-alignment hack necessary for gcc on Windows/i386. We will regret this in ten years (see previous change).
EDIT 2:
The dll became bad due to wrong options given to the configure script. The documentation is now updated.
You are correct. For an NxM float or double input you should allocate Nx(M/2+1) fftwf_complex or fftw_complex output.
The parameters are the input size. For example,
#include "fftw3.h"
#define Width 1024
#define Height 768
int main(){
float input[Width*Height];
fftwf_complex *fft = new fftwf_complex[((Width/2)+1)*Height];
fftwf_plan fplan = fftwf_plan_dft_r2c_2d(Height, Width,
(float*)input, fft, FFTW_ESTIMATE);
fftwf_execute(fplan);
fftwf_destroy_plan(fplan);
}
See Chapter 4.3 Basic Interface of the FFTW User Manual
When trying to compile a g++ project with the clang compiler, there is a strange error showing up.
Here is the snippet of the source file:
std::set<TTransportNetworkId> l_transportNetworkIds;
SelectionResultContainer l_searchResult = p_repo.rootMoc() / LnAny("LNBTS") / LnAny("LNMME");
BOOST_FOREACH(const SelectionResult & l_lnmmeSR, l_searchResult)
{
const MoLnmme & l_lnmme = l_lnmmeSR;
l_transportNetworkIds.insert(*l_lnmme.transportNwId);
}
The error message is:
conditional expression is ambiguous; 'rvalue_probe<Rrom::DataRep::SelectionResultContainer>' can be converted to 'Rrom::DataRep::SelectionResultContainer' and vice versa
BOOST_FOREACH(const SelectionResult & l_lnmmeSR, l_searchResult)
Conditions are:
The file compiles fine with gcc_4.3.2
clang in version 3.2 throws the above error
Already tried to include the latest boost library which results in the same error
My guess is that clang handles rvalue conditions differently than this gcc version does.
clang is supposed to be a drop-in-replacement for gcc, so how can one get rid of this error without touching the source file?
Are there any options in clang which somehow disables these kind of errors?!
UPDATE:
I could create an example source file, which you can reproduce for yourself:
#include <vector>
#include <boost/foreach.hpp>
class A : public std::vector<int>
{
public:
template <class T>
operator const T &() const;
};
void foo(){
A colA;
int b = 1;
BOOST_FOREACH(b, colA)
{
;
}
}
When compiled with clang 3.2 the above error is raised, with some additional insights to where exactly the error occurs:
error: conditional expression is ambiguous; 'rvalue_probe<A>' can be converted to 'A' and vice versa BOOST_FOREACH(b, colA)
expanded from macro 'BOOST_FOREACH' f (boost::foreach_detail_::auto_any_t BOOST_FOREACH_ID(_foreach_col) = BOOST_FOREACH_CONTAIN(COL))
expanded from macro 'BOOST_FOREACH_CONTAIN' BOOST_FOREACH_EVALUATE(COL)
expanded from macro 'BOOST_FOREACH_EVALUATE' (true ? boost::foreach_detail_::make_probe((COL), BOOST_FOREACH_ID(_foreach_is_rvalue)) : (COL))
This code is compiled without errors with gcc_4.7.2.
Any ideas why the two compilers behave differently?
I found the solution in this document, see http://www.boost.org/doc/libs/1_43_0/boost/foreach.hpp
Snippet:
// Some compilers do not correctly implement the lvalue/rvalue conversion
// rules of the ternary conditional operator.
# if defined(BOOST_FOREACH_NO_RVALUE_DETECTION)
So, when providing a -DBOOST_FOREACH_NO_RVALUE_DETECTION definition option to clang, the error disappears.
Still the question remains whether gcc or clang is right or wrong on this point.
I am trying to call cudppSort to sort a set of keys/values. I'm using the following code to set up the sort algorithm:
CUDPPConfiguration config;
config.op = CUDPP_ADD;
config.datatype = CUDPP_UINT;
config.algorithm = CUDPP_SORT_RADIX;
config.options = CUDPP_OPTION_KEY_VALUE_PAIRS | CUDPP_OPTION_FORWARD | CUDPP_OPTION_EXCLUSIVE;
CUDPPHandle planHandle;
CUDPPResult result = cudppPlan(&planHandle, config, number_points, 1, 0);
if (CUDPP_SUCCESS != result) {
printf("ERROR creating CUDPPPlan\n");
exit(-1);
}
The program exits, however on the line:
CUDPPResult result = cudppPlan(&planHandle, config, number_points, 1, 0);
and prints to stdout:
Cuda error: allocScanStorage in file 'c:/the/path/to/release1.1/cudpp/src/app/scan_app.cu' in line 279 : invalid configuration argument.
I looked at the line in scan_app.cu. It is,
CUT_CHECK_ERROR("allocScanStorage");
So apparently my configuration has an error that is causing the allocScanStorage to bomb out. There are only two calls to CUDA_SAFE_CALL in the function and I don't see a reason why either has anything to do with the configuration.
What is wrong with my configuration?
So that this doesn't sit around as an unanswered question (I'm not sure if this is the right SO etiquette but it seems like an answered question shouldn't sit around unanswered...), I'm copying the comment I made above here as an answer since it was the solution:
I figured this out (I'm still learning CUDA at the moment.) Because the error checking is asynchronous errors can show up in strange places if you don't check for them from time to time. My code had caused an error before I called cudppPlan but because I didn't check for errors the cudppPlan reported the error as if it was in cudppPlan.