Any example of clCreateProgramWithBinary use #540

davidcastells · 2022-02-28T10:19:26Z

davidcastells
Feb 28, 2022

I'm working with OpenCL on FPGAs, so I would like to load OpenCL programs from binary files.

All the examples I've seen (maybe I'm missing something) are creating programs from source. But I see that the code is somehow prepared to create programs from binary files.

Could you provide an example of how to do so?

BR

inducer · 2022-02-28T15:30:44Z

inducer
Feb 28, 2022
Maintainer

Following

https://documen.tician.de/pyopencl/runtime_program.html#pyopencl.Program

you can just do

prg = cl.Program(ctx, (device,), (binary,))

where binary is the bytes object containing the binary.

9 replies

inducer Feb 28, 2022
Maintainer

Run with gdb --args python ...?

davidcastells Mar 1, 2022
Author

It looks like it happens when the Altera (now Intel) code tries to free something. This does not happen in C/C++ code.
I still have to try with Xilinx bitstreams..

Thread 1 "python" received signal SIGSEGV, Segmentation fault.
__GI___libc_free (mem=0x7c046a007c047c03) at malloc.c:3113
3113    malloc.c: No such file or directory.
(gdb) bt
#0  __GI___libc_free (mem=0x7c046a007c047c03) at malloc.c:3113
#1  0x00007fffccc1fbd9 in l_load_binary_pkg () from /opt/intelFPGA/17.1/hld/host/linux64/lib/libalteracl.so
#2  0x00007fffccc2167a in clCreateProgramWithBinaryIntelFPGA () from /opt/intelFPGA/17.1/hld/host/linux64/lib/libalteracl.so
#3  0x00007ffff40c5bf3 in clCreateProgramWithBinary (context=0x158ae20, num_devices=1, device_list=0x15e2310, lengths=0x15e2750, binaries=0x15e2730,
    binary_status=<optimized out>, errcode_ret=0x7fffffffc3b8) at ocl_icd_loader_gen.c:2032
#4  0x00007ffff418f40f in pyopencl::create_program_with_binary(pyopencl::context&, pybind11::sequence, pybind11::sequence) ()
   from /home/dcr/.local/lib/python3.7/site-packages/pyopencl/_cl.cpython-37m-x86_64-linux-gnu.so
#5  0x00007ffff415d018 in void pybind11::cpp_function::initialize<void pybind11::detail::initimpl::factory<pyopencl_expose_part_2(pybind11::module_&)::{lambda(pyopencl::context&, pybind11::sequence, pybind11::sequence)#6}, pybind11::detail::void_type (*)(), pyopencl::program* (pyopencl::context&, pybind11::sequence, pybind11::sequence), pybind11::detail::void_type>::execute<pybind11::class_<pybind11::detail::void_type (*)()>, pybind11::arg, pybind11::class_<pybind11::detail::void_type (*)()>, pybind11::class_<pybind11::detail::void_type (*)()> >(pybind11::class_<pybind11::detail::void_type (*)()>&, pybind11::arg const&, pybind11::class_<pybind11::detail::void_type (*)()> const&, pybind11::class_<pybind11::detail::void_type (*)()> const&) &&::{lambda(pybind11::detail::value_and_holder&, pyopencl::context&, pybind11::sequence, pybind11::sequence)#1}, void, pybind11::detail::value_and_holder, pyopencl::context&, pybind11::sequence, pybind11::sequence, pybind11::name, pybind11::is_method, pybind11::sibling, pybind11::detail::is_new_style_constructor, pybind11::class_<pybind11::detail::void_type (*)()>, pybind11::class_<pybind11::detail::void_type (*)()>, pybind11::class_<pybind11::detail::void_type (*)()> >(pybind11::arg&&, void (*)(pybind11::detail::value_and_holder, pyopencl::context&, pybind11::sequence, pybind11::sequence), pybind11::name const&, pybind11::is_method const&, pybind11::sibling const&, pybind11::detail::is_new_style_constructor const&, pybind11::class_<pybind11::detail::void_type (*)()> const&, pybind11::class_<pybind11::detail::void_type (*)()> const&, pybind11::class_<pybind11::detail::void_type (*)()> const&)::{lambda(pybind11::detail::function_call&)#3}::_FUN ()
   from /home/dcr/.local/lib/python3.7/site-packages/pyopencl/_cl.cpython-37m-x86_64-linux-gnu.so
#6  0x00007ffff411bd0f in pybind11::cpp_function::dispatcher(_object*, _object*, _object*) ()
   from /home/dcr/.local/lib/python3.7/site-packages/pyopencl/_cl.cpython-37m-x86_64-linux-gnu.so
#7  0x00000000005bc266 in _PyMethodDef_RawFastCallDict () at Objects/call.c:515
#8  0x00000000004c310a in _PyCFunction_FastCallDict (kwargs=0x0, nargs=4, args=0x7fffffffc7f0, func=0x7ffff4086e10) at Objects/call.c:582
#9  _PyObject_FastCallDict (kwargs=0x0, nargs=4, args=0x7fffffffc7f0, callable=0x7ffff4086e10) at Objects/call.c:101
#10 _PyObject_Call_Prepend () at Objects/call.c:904
#11 0x00000000004c3786 in PyObject_Call () at Objects/call.c:245
#12 0x000000000043bd49 in slot_tp_init (self=0x7ffff3fde150, args=0x7ffff4086480, kwds=0x0) at Objects/typeobject.c:6613
#13 0x0000000000501d19 in type_call () at Objects/typeobject.c:949
#14 0x00007ffff411a0cb in pybind11_meta_call () from /home/dcr/.local/lib/python3.7/site-packages/pyopencl/_cl.cpython-37m-x86_64-linux-gnu.so
#15 0x00000000004c17fa in _PyObject_FastCallKeywords () at Objects/call.c:199
#16 0x0000000000538e46 in call_function (kwnames=0x0, oparg=<optimized out>, pp_stack=<synthetic pointer>) at Python/ceval.c:4619
#17 _PyEval_EvalFrameDefault () at Python/ceval.c:3093
#18 0x0000000000532bb5 in PyEval_EvalFrameEx (throwflag=0, f=0x140a328) at Python/ceval.c:547
#19 _PyEval_EvalCodeWithName () at Python/ceval.c:3930
#20 0x00000000004c3464 in _PyFunction_FastCallDict (kwargs=0x0, nargs=4, args=0x7fffffffcc10, func=0x7ffff4069158) at Objects/call.c:376
#21 _PyObject_FastCallDict (kwargs=0x0, nargs=4, args=0x7fffffffcc10, callable=0x7ffff4069158) at Objects/call.c:98
#22 _PyObject_Call_Prepend () at Objects/call.c:904
#23 0x0000000000504da7 in slot_tp_init () at Objects/typeobject.c:6610
#24 0x0000000000501d19 in type_call () at Objects/typeobject.c:949
#25 0x00000000004c17fa in _PyObject_FastCallKeywords () at Objects/call.c:199
#26 0x0000000000538e46 in call_function (kwnames=0x0, oparg=<optimized out>, pp_stack=<synthetic pointer>) at Python/ceval.c:4619
#27 _PyEval_EvalFrameDefault () at Python/ceval.c:3093
#28 0x0000000000532bb5 in PyEval_EvalFrameEx (throwflag=0, f=0x15f0c78) at Python/ceval.c:547
#29 _PyEval_EvalCodeWithName () at Python/ceval.c:3930
#30 0x00000000004c2ac3 in _PyFunction_FastCallKeywords () at Objects/call.c:433
#31 0x00000000005383bc in call_function (kwnames=0x0, oparg=<optimized out>, pp_stack=<synthetic pointer>) at Python/ceval.c:4616
#32 _PyEval_EvalFrameDefault () at Python/ceval.c:3093
#33 0x0000000000532bb5 in PyEval_EvalFrameEx (throwflag=0, f=0x992ab8) at Python/ceval.c:547
#34 _PyEval_EvalCodeWithName () at Python/ceval.c:3930
---Type <return> to continue, or q <return> to quit---
#35 0x0000000000532a63 in PyEval_EvalCodeEx (closure=0x0, kwdefs=0x0, defcount=0, defs=0x0, kwcount=0, kws=0x0, argcount=0, args=0x0, locals=locals@entry=0x7ffff7ecc1f8,
    globals=globals@entry=0x7ffff7ecc1f8, _co=_co@entry=0x7ffff7f05660) at Python/ceval.c:3959
#36 PyEval_EvalCode (co=co@entry=0x7ffff7f05660, globals=globals@entry=0x7ffff7ecc1f8, locals=locals@entry=0x7ffff7ecc1f8) at Python/ceval.c:524
#37 0x00000000005de0f2 in run_mod () at Python/pythonrun.c:1035
#38 0x00000000005de52a in PyRun_FileExFlags () at Python/pythonrun.c:988
#39 0x00000000005de317 in PyRun_SimpleFileExFlags () at Python/pythonrun.c:429
#40 0x00000000005b70f1 in pymain_run_file (p_cf=0x7fffffffd2b0, filename=0x94f840 L"perf_close3d_oclfpga_closingperpixel.py", fp=0x997360) at Modules/main.c:427
#41 pymain_run_filename (cf=0x7fffffffd2b0, pymain=0x7fffffffd3c0) at Modules/main.c:1627
#42 pymain_run_python (pymain=0x7fffffffd3c0) at Modules/main.c:2876
#43 pymain_main () at Modules/main.c:3037
#44 0x00000000005b6dba in _Py_UnixMain () at Modules/main.c:3072
#45 0x00007ffff703fbf7 in __libc_start_main (main=0x4bd810 <main>, argc=2, argv=0x7fffffffd518, init=<optimized out>, fini=<optimized out>, rtld_fini=<optimized out>,
    stack_end=0x7fffffffd508) at ../csu/libc-start.c:310
#46 0x00000000005b6ca9 in _start ()

inducer Mar 1, 2022
Maintainer

This looks like an issue with the Altera ICD, not sure that PyOpenCL could do to avoid this issue.

davidcastells Mar 13, 2022
Author

I doesn't look like.
I've discovered that the error is different when using lists or tuples

This code
program = cl.Program(context, [device], [binary])
....raises the free error and crash the app

and this code
program = cl.Program(context, (device,), (binary,))
...raises and INVALID_BINARY error exiting in a controlled way

Still a similar code in (C/C++) works perfectly

I get this behaviour in pyopencl versions <= 2020.1

pyopencl versions >= 2020.2 don't work with my FPGA because it seems that pyopencl is not supporting OpenCL versions < 2.1

ImportError: /opt/intelFPGA/17.1/hld/host/linux64/lib/libOpenCL.so.1: version `OPENCL_2.1' not found (required by /home/dcr/.local/lib/python3.7/site-packages/pyopencl/_cl.cpython-37m-x86_64-linux-gnu.so)

Is this so? Or otherwise, can I fix it somehow?

davidcastells Mar 13, 2022
Author

OK. I solved the issue 🎉.
Last pyopencl version is working.

And the solution is 🥁🥁🥁.... removing a "import matplotlib.pyplot" that I had

Not joking, apparently it makes pyopencl on my fpga crash. Don't ask my why, it's really wierd

davidcastells · 2022-03-01T09:15:54Z

davidcastells
Mar 1, 2022
Author

Let me provide more info.
I've been trying to trigger the same error from a C++ test program without success.

The program prints SUCCESS when done right, and raises a segmentation fault when I introduce invalid parameters, but never in free from libc (as we get when called from PyOpenCL).

I attach the small code snippet with comments about the tests I did.

Any idea ?

 cl_device_id devices[]={did};
 size_t fileLen;
 unsigned char* pFile = loadBinaryFile("close_nd.aocx", &fileLen);

 const unsigned char* binaries[]={pFile};

 size_t binary_lengths[]={fileLen};
 cl_int binary_status[1];   
 // context = (cl_context) 3;  // THIS CAUSES A SEGMENTATION FAULT before entering the altera library
 // cl_int* pErr = (cl_int*) 5; // THIS CAUSES A SEGMENTATION FAULT in the altera library, but not in free
 // cl_int* binary_status = (cl_int*) 4; // THIS CAUSES A SEGMENTATION FAULT in the altera library, but not in free
 // size_t* binary_lengths = (size_t*) 4; // THIS CAUSES A SEGMENTATION FAULT in the altera library, but not in free
 // const unsigned char* binaries[] = {(unsigned char*) 64};  // THIS CAUSES A SEGMENTATION FAULT in the altera library when trying to copy (not in free)
 //  const unsigned char** binaries = (const unsigned char**) 4;  // THIS CAUSES A SEGMENTATION FAULT in the altera library, but not in free

 cl_program program = clCreateProgramWithBinary(context, 1, devices, binary_lengths, binaries, binary_status, &err);
 assert(err == CL_SUCCESS);
 printf("SUCCESS\n");

4 replies

yves-surrel Mar 13, 2022

It happens that I just came into that problem.

On my Mac MBP, the following code:

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sun Mar 13 09:58:27 2022

@author: yves
"""

import os
import pyopencl as cl

# Kernel function
src = """
__kernel void test(
    __global int *a)
{
    a[get_global_id(0)] = 0;
}
"""

platforms = cl.get_platforms()

# Save binaries
for i_platform, platform in enumerate(platforms):
    devices = platform.get_devices()
    for i_device, device in enumerate(devices):
        ctx = cl.Context([device])
        prg = cl.Program(ctx, src).build()
        binary, = prg.get_info(cl.program_info.BINARIES)
        print(prg.get_info(cl.program_info.BINARY_SIZES)[0])
        filename = f'test_{i_platform}_{i_device}.bin'
        if os.path.exists(filename):
            os.remove(filename)
        with open(filename, 'xb') as bin:
            bin.write(binary)
            

# Read binaries and create programs
for i_platform, platform in enumerate(platforms):
    devices = platform.get_devices()
    for i_device, device in enumerate(devices):
        ctx = cl.Context([device])
        with open(f'test_{i_platform}_{i_device}.bin', 'rb') as bin:
            binary = bin.read()
            print(len(binary))
            prg = cl.Program(ctx, [device], [binary])
            print(prg.get_info(cl.program_info.BINARY_SIZES)[0])
            try:
                print(prg.all_kernels())
            except cl.LogicError as e:
                print(e)

returns

223
1796
13659
223
0
clCreateKernelsInProgram failed: INVALID_PROGRAM_EXECUTABLE
1796
0
clCreateKernelsInProgram failed: INVALID_PROGRAM_EXECUTABLE
13659
0
clCreateKernelsInProgram failed: INVALID_PROGRAM_EXECUTABLE

I should be missing something...

davidcastells Mar 13, 2022
Author

In your example you compile and save the binary file from pyopencl, and then program the binary file from pyopencl as well. As it fails, we don't know wether the error is from the compilation phase or the programming from binary phase.

Can you separe building from programming ? and test that you are able to program the binary from a C/C++ executable?
By the way, can you try if using tuples instead of lists make any difference?

prg=cl.Program(ctx,(device,),(binary,))

yves-surrel Mar 13, 2022

It's the same with prg = cl.Program(ctx, (device,), (binary,)) (contrary would have been surprising...)

I do not understand what you mean by 'separe building from programming'. What kind of change in the source do you expect?

My example should be runnable on your computer (Minimal Working Example). How does it behave?

yves-surrel Mar 13, 2022

...and I am not a C++ programmer

yves-surrel · 2022-03-13T11:26:40Z

yves-surrel
Mar 13, 2022

An even smaller working example (without writing files):

import os
import pyopencl as cl

# Kernel function
src = """
__kernel void test(
    __global int *a)
{
    a[get_global_id(0)] = 0;
}
"""

platforms = cl.get_platforms()

for platform in platforms:
    devices = platform.get_devices()
    for i_device, device in enumerate(devices):
        ctx = cl.Context([device])
        prg = cl.Program(ctx, src).build()
        binary, = prg.get_info(cl.program_info.BINARIES)
        # Create Program from binary
        prg1 = cl.Program(ctx, (device,), (binary,))
        try:
            print(prg1.all_kernels())
        except cl.LogicError as e:
            print(e)

0 replies

yves-surrel · 2022-03-13T15:26:24Z

yves-surrel
Mar 13, 2022

On my BootCamp Windows, it simply crashes python :-(

0 replies

yves-surrel · 2022-04-04T20:28:34Z

yves-surrel
Apr 4, 2022

I have found somewhere in StackOverflow: " if you want to save a compiled kernel, save it after a call to clCompileProgram method. On #Mojave you cannot build a program from binary if that binary was previously built by clBuildProgram". Seems true.
If I replace buildby compile, it works:

import os
import pyopencl as cl

# Kernel function
src = """
__kernel void test(
    __global int *a)
{
    a[get_global_id(0)] = 0;
}
"""

platforms = cl.get_platforms()

for platform in platforms:
    devices = platform.get_devices()
    for i_device, device in enumerate(devices):
        ctx = cl.Context([device])
        prg = cl.Program(ctx, src).compile()
        sizes, = prg.get_info(cl.program_info.BINARY_SIZES)
        binary, = prg.get_info(cl.program_info.BINARIES)
        # Create Program from binary
        prg1 = cl.Program(ctx, (device,), (binary,))
        prg1.build()
        for knl in prg1.all_kernels():
            print(knl.function_name)

prints

test
test
test
/Volumes/FAT/Developpement/Python/core/untitled6.py:27: UserWarning: Pre-build attribute access defeats compiler caching.
  prg = cl.Program(ctx, src).compile()

I do not know how to suppress the warning (and the relevance of the warning either).

1 reply

inducer Apr 4, 2022
Maintainer

prg.get_info(cl.program_info.BINARY_SIZES) can be prg.binary_sizes. (etc.)

I do not know how to suppress the warning (and the relevance of the warning either).

The warnings mean that the PyOpenCL compiler cache won't be used. If you're not waiting longer than you're willing for things to compile, then it's safe to ignore.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Any example of clCreateProgramWithBinary use #540

{{title}}

Replies: 5 comments 14 replies

{{title}}

{{title}}

{{title}}

{{editor}}'s edit

{{editor}}'s edit

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

{{title}}

Select a reply

Any example of clCreateProgramWithBinary use #540

davidcastells Feb 28, 2022

Replies: 5 comments · 14 replies

inducer Feb 28, 2022 Maintainer

inducer Feb 28, 2022 Maintainer

davidcastells Mar 1, 2022 Author

inducer Mar 1, 2022 Maintainer

davidcastells Mar 13, 2022 Author

davidcastells Mar 13, 2022 Author

davidcastells Mar 1, 2022 Author

yves-surrel Mar 13, 2022

davidcastells Mar 13, 2022 Author

yves-surrel Mar 13, 2022

yves-surrel Mar 13, 2022

yves-surrel Mar 13, 2022

yves-surrel Mar 13, 2022

yves-surrel Apr 4, 2022

inducer Apr 4, 2022 Maintainer

davidcastells
Feb 28, 2022

Replies: 5 comments 14 replies

inducer
Feb 28, 2022
Maintainer

inducer Feb 28, 2022
Maintainer

davidcastells Mar 1, 2022
Author

inducer Mar 1, 2022
Maintainer

davidcastells Mar 13, 2022
Author

davidcastells Mar 13, 2022
Author

davidcastells
Mar 1, 2022
Author

davidcastells Mar 13, 2022
Author

yves-surrel
Mar 13, 2022

yves-surrel
Mar 13, 2022

yves-surrel
Apr 4, 2022

inducer Apr 4, 2022
Maintainer