Remove CUDA 11 from code, docs, etc.

[jakirkham]

Description

CUDA 11 is being removed in RAPIDS 25.08 as noted in the RSN & issue below.

CI and package builds for CUDA 11 have already been dropped.

However references in the code, docs, etc. are still present.

This issue tracks removing all remaining CUDA 11 references.

References:

• RSN 48
• Drop CUDA 11 build-planning#184

[jakirkham]

Some old bits of code:

cudf/cpp/src/jit/runtime_support.cpp

Lines 24 to 32 in 68f3185

	bool is_runtime_jit_supported()
	{
	// We require a CUDA NVRTC of 11.5+ to do runtime jit
	// as we need support for __int128
	int runtime = 0;
	CUDF_CUDA_TRY(cudaRuntimeGetVersion(&runtime));
	return runtime >= 11050;
	}

cudf/python/cudf/cudf/utils/gpu_utils.py

Lines 111 to 119 in 68f3185

	if cuda_runtime_version < 11000:
	# Require CUDA Runtime version 11.0 or greater.
	major_version = cuda_runtime_version // 1000
	minor_version = (cuda_runtime_version % 1000) // 10
	raise UnsupportedCUDAError(
	"Detected CUDA Runtime version is "
	f"{major_version}.{minor_version}. "
	"Please update your CUDA Runtime to 11.0 or above."
	)

cudf/cpp/include/cudf/detail/device_scalar.hpp

Lines 38 to 42 in 68f3185

	// Implementation is the same as what compiler should generate
	// Could not use default move constructor as 11.8 compiler fails to generate it
	#ifdef __CUDACC__
	#pragma nv_exec_check_disable
	#endif

cudf/cpp/include/cudf/utilities/bit.hpp

Lines 32 to 39 in 68f3185

	// @cond
	// Work around a bug in NVRTC that fails to compile assert() in constexpr
	// functions (fixed after CUDA 11.0)
	#if defined __GNUC__
	#define LIKELY(EXPR) __builtin_expect(!!(EXPR), 1)
	#else
	#define LIKELY(EXPR) (!!(EXPR))
	#endif

cudf/cpp/src/stream_compaction/filter/filter.cu

Line 186 in 68f3185

CUDF_EXPECTS(is_runtime_jit_supported(), "Runtime JIT is only supported on CUDA Runtime 11.5+");

cudf/cpp/include/cudf/fixed_point/fixed_point.hpp

Lines 82 to 91 in 68f3185

	CUDF_HOST_DEVICE constexpr inline scale_type min(scale_type const& a, scale_type const& b)
	{
	// TODO This is a temporary workaround because <cuda/std/functional> is not self-contained when
	// built with NVRTC 11.8. Replace this with cuda::std::min once the underlying issue is resolved.
	#ifdef __CUDA_ARCH__
	return scale_type{min(static_cast<int>(a), static_cast<int>(b))};
	#else
	return std::min(a, b);
	#endif
	}

[jakirkham]

Also some old docs:

cudf/python/custreamz/README.md

Lines 49 to 53 in 68f3185

	### CUDA/GPU requirements
	* CUDA 11.0+
	* NVIDIA driver 450.80.02+
	* Volta architecture or better (Compute Capability >=7.0)

cudf/docs/cudf/source/user_guide/io/io.md

Lines 88 to 93 in 68f3185

	Many IO APIs can use Magnum IO GPUDirect Storage (GDS) library to optimize
	IO operations. GDS enables a direct data path for direct memory access
	(DMA) transfers between GPU memory and storage, which avoids a bounce
	buffer through the CPU. The SDK is available for download
	[here](https://developer.nvidia.com/gpudirect-storage). GDS is also
	included in CUDA Toolkit 11.4 and higher.

[galipremsagar]

With the last PR: #19386 now merged, I think we are good to close this issue @jakirkham ?

[jakirkham]

Thanks Prem! 🙏

Am still seeing a few CUDA 11 references

cudf/cpp/src/jit/runtime_support.cpp

Lines 24 to 32 in 55ad793

	bool is_runtime_jit_supported()
	{
	// We require a CUDA NVRTC of 11.5+ to do runtime jit
	// as we need support for __int128
	int runtime = 0;
	CUDF_CUDA_TRY(cudaRuntimeGetVersion(&runtime));
	return runtime >= 11050;
	}

cudf/cpp/src/transform/transform.cpp

Line 265 in 55ad793

CUDF_EXPECTS(is_runtime_jit_supported(), "Runtime JIT is only supported on CUDA Runtime 11.5+");

cudf/cpp/src/stream_compaction/filter/filter.cu

Line 186 in 55ad793

CUDF_EXPECTS(is_runtime_jit_supported(), "Runtime JIT is only supported on CUDA Runtime 11.5+");

Given these should always be true, think we can just drop all of these