CUDA: Understanding PTX Information

Question

I do not find much useful information about the PTX information --ptxas-options=-v I found the 2008 NVCC pdf, which has a small commercial, but no details.
1) What does 64 bytes cmem[0], 12 bytes cmem[16] ? I understand that this refers to read-only memory. I do not use constant mem in the code, so this should come from the compiler. (What is included in RO mem?)
2) What does 49152+0 bytes smem ? Yes, this is shared memory, but what do these two mean?
3) Is there a document that will help me with this? (What is called?)
4) Where can I find a document that will explain the * .ptx file? (I would like to be able to read / understand the cuda assy code.)

Answer 1

• cmem discussed here . In your case, this means that 64 bytes are used to pass arguments to the kernel, and 12 bytes are occupied by the constants generated by the compiler.

• In the case of smem first number is the amount of data in your code request, and the second number ( 0 ) indicates how much memory is used for system purposes.

• I do not know any official information about the detailed output format of ptxas . For example. in CUDA's Employment Calculator, they simply say to summarize the values ​​for smem without any explanation.

• There are several PTX documents on the nVidia website. The most fundamental is PTX: a parallel implementation of the ISA stream version 3.0 .

Answer 2

See “Other Uses for NVCC .” They note that the permanent distribution of banks depends on the profile .

In the PTX manual , they say that besides 64KB of read-only memory, they also had 10 banks for read-only memory. A driver can allocate and initialize persistent buffers in these regions and pass pointers to buffers as parameters of the kernel function.

I think this profile set for nvcc will take care of which constants fall into which memory. In any case, we need not worry if each cmem [n] read-only memory is less than 64 KB , since each bank is 64 KB in size and is common to all threads in the grid.