CUDA Support for GNU Radio using the custom buffer changes introduced in GR 3.10. Custom buffers for CUDA-enabled hardware are provided that can be included in any OOT. This allows the work() or general_work() function of a block to have data provided in the input- and output- buffers that has already been transitioned to device memory so that data movement does not have to take place in the work() function itself.
This OOT is adapted/copied from the gr-cuda_buffer OOT work by Black Lynx, Inc. (https://github.com/BlackLynx-Inc/gr-cuda_buffer) and relies on the custom buffer feature developed by David Sorber and documented here: https://wiki.gnuradio.org/index.php/CustomBuffers
- NVIDIA CUDA supported GPU
- Installation of CUDA
- GNU Radio >= 3.10.0.0
This OOT can be included in your OOT as follows
project(gr-myoot CXX C CUDA)- Find the exported CMake from gr-cuda as follows
find_package(gnuradio-cuda REQUIRED)- Create a library to hold the CUDA specific code:
add_library(gnuradio-myoot-cu STATIC
my_kernel.cu
)
set_target_properties(gnuradio-myoot-cu PROPERTIES
POSITION_INDEPENDENT_CODE ON
CUDA_VISIBILITY_PRESET "hidden"
CUDA_SEPARABLE_COMPILATION ON
)
# Make a fat binary for supported architectures
# Compile for supported archs; CMake workaround, see: https://stackoverflow.com/a/54505975/73878
target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_60,code=sm_60>")
target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_61,code=sm_61>")
target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_62,code=sm_62>")
target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_70,code=sm_70>")
target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_72,code=sm_72>")
target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_75,code=sm_75>")
# target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_80,code=sm_80>")
# target_compile_options(gnuradio-myoot-cu PRIVATE "$<$<COMPILE_LANGUAGE:CUDA>:SHELL:-gencode arch=compute_86,code=sm_86>")- Link to gr-cuda
The target_link_libraries statement should be changed to specify public linkage
target_link_libraries(gnuradio-myoot PUBLIC gnuradio::gnuradio-runtime gnuradio-cuda)- Link to the custom CUDA code
target_link_libraries(gnuradio-myoot PRIVATE gnuradio-myoot-cu)- Create a
.cufile that will hold the CUDA kernel and wrapper and be compiled withnvccand linked to the block
See multiply_const.cu for an example
- Use
gr_modtoolto create a block in your OOT
See multiply_const.h, multiply_const_impl.{cc,h}
-
Set the io_signature to specify the custom buffer
-
Call the kernel wrapper inside the
work()function