Add a helper function to map PTX ISA versions to the corresponding CTK versions

[leofang]

We want to bite the bullet and encode this table as a helper function in cuda.bindings:
https://docs.nvidia.com/cuda/parallel-thread-execution/#release-notes-ptx-release-history
Ex: PTX ISA 8.8 -> CTK 12.9. This is important for us to determine if a PTX is loadable by the driver or not.

[rwgk]

Copy-pasting from comment on meeting doc:

That implementation essentially exists already, although currently only as a unit test:

https://github.com/NVIDIA/cuda-python/blob/ed884c97a7d5213d520ae9bd9415409b22c10e44/cuda_bindings/tests/test_nvvm.py#L22-L37

It's probably an easy afternoon project to convert that to simple API, e.g. cuda.bindings.nvvm.get_isa_version()

(Going from there to all the other information in https://docs.nvidia.com/cuda/parallel-thread-execution/#release-notes-ptx-release-history seems even easier.)

For completeness, the input first:

target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-i128:128:128-f32:32:32-f64:64:64-v16:16:16-v32:32:32-v64:64:64-v128:128:128-n16:32:64"

target triple = "nvptx64-nvidia-cuda"

define void @kernel() {
entry:
ret void
}

!nvvm.annotations = !{!0}
!0 = !{void ()* @kernel, !"kernel", i32 1}

!nvvmir.version = !{!1}
!1 = !{i32 2, i32 0, i32 3, i32 0}

This is the output:

//
// Generated by NVIDIA NVVM Compiler
//
// Compiler Build ID: CL-35813241
// Cuda compilation tools, release 12.9, V12.9.41
// Based on NVVM 7.0.1
//

.version 8.8
.target sm_52
.address_size 64

// .globl kernel

.visible .entry kernel()
{



ret;

}

---

$ git diff
diff --git a/cuda_bindings/tests/test_nvvm.py b/cuda_bindings/tests/test_nvvm.py
index db8481a9..3988e12e 100644
--- a/cuda_bindings/tests/test_nvvm.py
+++ b/cuda_bindings/tests/test_nvvm.py
@@ -263,6 +263,9 @@ def test_get_buffer_empty(get_size, get_buffer):
 @pytest.mark.parametrize("options", [[], ["-opt=0"], ["-opt=3", "-g"]])
 def test_compile_program_with_minimal_nvvm_ir(minimal_nvvmir, options):
     with nvvm_program() as prog:
+        print("\nINPUT")
+        print(minimal_nvvmir.decode())
+        print()
         nvvm.add_module_to_program(prog, minimal_nvvmir, len(minimal_nvvmir), "FileNameHere.ll")
         try:
             nvvm.compile_program(prog, len(options), options)
@@ -278,6 +281,9 @@ def test_compile_program_with_minimal_nvvm_ir(minimal_nvvmir, options):
         buffer = bytearray(result_size)
         nvvm.get_compiled_result(prog, buffer)
         assert ".visible .entry kernel()" in buffer.decode()
+        print("\nOUTPUT")
+        print(buffer.decode())
+        print()

LD_LIBRARY_PATH=/usr/local/cuda/nvvm/lib64 pytest -s -vv test_nvvm.py -k 'test_compile_program_with_minimal_nvvm_ir[txt-options0]'

[rwgk]

Keeping track of a related URL provided by @bernhardmgruber:

https://github.com/NVIDIA/cccl/blob/502e26f94f5dc3e7ceb2c60c3d3121d255b997be/libcudacxx/include/cuda/std/__cccl/ptx_isa.h#L39

[leofang]

Yes something like what the ptx_isa.h header does, but reversed (we want PTX -> CUDA not CUDA -> PTX). Also we can't use that header as-is today at build time, as we use only gcc and no nvcc/clang.

[leofang]

That implementation essentially exists already, although currently only as a unit test:

I think this should live in numba-cuda (as part of NVIDIA/numba-cuda#262), not cuda.bindings?

[rwgk]

It would be more generally accessible in cuda.bindings.

I'm not sure though how "pure" you want to keep cuda.bindings. The functionality is somewhere in a gray area, between cuda.bindings and cuda.core (similar to the path_finder).

IMHO: I'd add it in cuda.bindings.

[leofang]

I certainly think cuda.bindings is the right place for hosting the PTX ISA table, but not so sure about the bits for extracting ISA version from an NVVM IR. I feel it is the responsibility of whoever consuming NVVM IRs as inputs. That said, very soon one of them will be cuda.core (#452), so perhaps this is something that @tpn can reuse there.