nvcc & PTX: its internal operation and how to generate it

Go to any of the CUDA examples:

https://docs.nvidia.com/cuda/cuda-samples/index.html

And edit the Makefile to modify two things:   NVCCFLAGS and gpuarch codes.

First is the NVCCFLAGS:  append a "-ptx" to the end. 

And the SMS code, which enumerate through all the different GPU architecture, just modify it to a single GPU architecture (for eg, 30 is used here):

And the details is explained here:   (https://docs.nvidia.com/cuda/cuda-samples/index.html#getting-cuda-samples)  

And then "make" will generate:

"/home/tthtlc/cuda-10.0"/bin/nvcc -ccbin g++ -I../../common/inc  -m64 -ptx     -gencode arch=compute_35,code=sm_35 -gencode arch=compute_35,code=compute_35 -o dxtc.o -c dxtc.cu"/home/tthtlc/cuda-10.0"/bin/nvcc -ccbin g++   -m64 -ptx       -gencode arch=compute_35,code=sm_35 -gencode arch=compute_35,code=compute_35 -o dxtc dxtc.o
mkdir -p ../../bin/x86_64/linux/release
cp dxtc ../../bin/x86_64/linux/release

Then edit the dxtc.o file, which consist of all the PTX instructions:

https://pastebin.com/McnJ4vYx

https://developer.download.nvidia.com/compute/DevZone/docs/html/C/doc/nvcc.pdf 

https://docs.nvidia.com/cuda/parallel-thread-execution/index.html 

https://en.wikipedia.org/wiki/Parallel_Thread_Execution 

and the latest ISA specs is here:

https://docs.nvidia.com/cuda/pdf/ptx_isa_6.4.pdf

For example:

 https://docs.nvidia.com/cuda/parallel-thread-execution/index.html 

https://devtalk.nvidia.com/default/topic/551214/teaching-and-curriculum-support/gpu-assembly/

https://devtalk.nvidia.com/default/topic/492195/problem-generating-ptx-files/

PTX have a unique features in its memory model:

https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#memory-hierarchy