Merge pull request #7894 from obilaniu/cudaupdates

Update unstable CUDA Module to support newest toolkits and drivers.

3 files changed, 169 insertions, 58 deletions

diff --git a/docs/markdown/Cuda-module.md b/docs/markdown/Cuda-module.md
index f161eac..2f8cb02 100644
--- a/docs/markdown/Cuda-module.md
+++ b/docs/markdown/Cuda-module.md
@@ -35,14 +35,14 @@ It offers several useful functions that are enumerated below.
 _Since: 0.50.0_
 
 ``` meson
-cuda.nvcc_arch_flags(nvcc_or_version, ...,
+cuda.nvcc_arch_flags(cuda_version_string, ...,
                      detected: string_or_array)
 ```
 
 Returns a list of `-gencode` flags that should be passed to `cuda_args:` in
 order to compile a "fat binary" for the architectures/compute capabilities
 enumerated in the positional argument(s). The flags shall be acceptable to
-the NVCC compiler object `nvcc_or_version`, or its version string.
+an NVCC with CUDA Toolkit version string `cuda_version_string`.
 
 A set of architectures and/or compute capabilities may be specified by:
 
@@ -71,14 +71,6 @@ mixed with architecture names or compute capabilities. Their interpretation is:
 | `'Common'`        | Relatively common CCs supported by given NVCC compiler. Generally excludes Tegra and Tesla devices. |
 | `'Auto'`          | The CCs provided by the `detected:` keyword, filtered for support by given NVCC compiler. |
 
-As a special case, when `nvcc_arch_flags()` is invoked with
-
-- an NVCC `compiler` object `nvcc`,
-- `'Auto'` mode and
-- no `detected:` keyword,
-
-Meson uses `nvcc`'s architecture auto-detection results.
-
 The supported architecture names and their corresponding compute capabilities
 are:
 
@@ -95,6 +87,7 @@ are:
 | `'Volta'`         | 7.0                |
 | `'Xavier'`        | 7.2                |
 | `'Turing'`        | 7.5                |
+| `'Ampere'`        | 8.0, 8.6           |
 
 
 Examples:
@@ -152,7 +145,7 @@ function `CUDA_SELECT_NVCC_ARCH_FLAGS(out_variable, [list of CUDA compute archit
 _Since: 0.50.0_
 
 ``` meson
-cuda.nvcc_arch_readable(nvcc_or_version, ...,
+cuda.nvcc_arch_readable(cuda_version_string, ...,
                         detected: string_or_array)
 ```
 
@@ -162,7 +155,7 @@ architectures that will be compiled for. The output of this function is solely
 intended for informative message printing.
 
     archs    = '3.0 3.5 5.0+PTX'
-    readable = cuda.nvcc_arch_readable(nvcc, archs)
+    readable = cuda.nvcc_arch_readable('10.0', archs)
     message('Building for architectures ' + ' '.join(readable))
 
 This will print
@@ -178,11 +171,11 @@ _Note:_ This function is intended to closely replicate CMake's FindCUDA module f
 _Since: 0.50.0_
 
 ``` meson
-cuda.min_driver_version(nvcc_or_version)
+cuda.min_driver_version(cuda_version_string)
 ```
 
 Returns the minimum NVIDIA proprietary driver version required, on the host
-system, by kernels compiled with the given NVCC compiler or its version string.
+system, by kernels compiled with a CUDA Toolkit with the given version string.
 
 The output of this function is generally intended for informative message
 printing, but could be used for assertions or to conditionally enable
diff --git a/mesonbuild/modules/unstable_cuda.py b/mesonbuild/modules/unstable_cuda.py
index 0d693c3..0f9d681 100644
--- a/mesonbuild/modules/unstable_cuda.py
+++ b/mesonbuild/modules/unstable_cuda.py
@@ -34,16 +34,19 @@ class CudaModule(ExtensionModule):
     @noKwargs
     def min_driver_version(self, state, args, kwargs):
         argerror = InvalidArguments('min_driver_version must have exactly one positional argument: ' +
-                                    'an NVCC compiler object, or its version string.')
+                                    'a CUDA Toolkit version string. Beware that, since CUDA 11.0, ' +
+                                    'the CUDA Toolkit\'s components (including NVCC) are versioned ' +
+                                    'independently from each other (and the CUDA Toolkit as a whole).')
 
-        if len(args) != 1:
+        if len(args) != 1 or not isinstance(args[0], str):
             raise argerror
-        else:
-            cuda_version = self._version_from_compiler(args[0])
-            if cuda_version == 'unknown':
-                raise argerror
 
+        cuda_version = args[0]
         driver_version_table = [
+            {'cuda_version': '>=11.1.0',   'windows': '456.38', 'linux': '455.23'},
+            {'cuda_version': '>=11.0.3',   'windows': '451.82', 'linux': '450.51.06'},
+            {'cuda_version': '>=11.0.2',   'windows': '451.48', 'linux': '450.51.05'},
+            {'cuda_version': '>=11.0.1',   'windows': '451.22', 'linux': '450.36.06'},
             {'cuda_version': '>=10.2.89',  'windows': '441.22', 'linux': '440.33'},
             {'cuda_version': '>=10.1.105', 'windows': '418.96', 'linux': '418.39'},
             {'cuda_version': '>=10.0.130', 'windows': '411.31', 'linux': '410.48'},
@@ -128,19 +131,45 @@ class CudaModule(ExtensionModule):
 
         return cuda_version, arch_list, detected
 
+    def _filter_cuda_arch_list(self, cuda_arch_list, lo=None, hi=None, saturate=None):
+        """
+        Filter CUDA arch list (no codenames) for >= low and < hi architecture
+        bounds, and deduplicate.
+        If saturate is provided, architectures >= hi are replaced with saturate.
+        """
+
+        filtered_cuda_arch_list = []
+        for arch in cuda_arch_list:
+            if arch:
+                if lo and version_compare(arch, '<' + lo):
+                    continue
+                if hi and version_compare(arch, '>=' + hi):
+                    if not saturate:
+                        continue
+                    arch = saturate
+                if arch not in filtered_cuda_arch_list:
+                    filtered_cuda_arch_list.append(arch)
+        return filtered_cuda_arch_list
+
     def _nvcc_arch_flags(self, cuda_version, cuda_arch_list='Auto', detected=''):
         """
-        Using the CUDA Toolkit version (the NVCC version) and the target
-        architectures, compute the NVCC architecture flags.
+        Using the CUDA Toolkit version and the target architectures, compute
+        the NVCC architecture flags.
         """
 
-        cuda_known_gpu_architectures  = ['Fermi', 'Kepler', 'Maxwell']  # noqa: E221
-        cuda_common_gpu_architectures = ['3.0', '3.5', '5.0']           # noqa: E221
-        cuda_limit_gpu_architecture   = None                            # noqa: E221
-        cuda_all_gpu_architectures    = ['3.0', '3.2', '3.5', '5.0']    # noqa: E221
+        # Replicates much of the logic of
+        #     https://github.com/Kitware/CMake/blob/master/Modules/FindCUDA/select_compute_arch.cmake
+        # except that a bug with cuda_arch_list="All" is worked around by
+        # tracking both lower and upper limits on GPU architectures.
+
+        cuda_known_gpu_architectures   = ['Fermi', 'Kepler', 'Maxwell']  # noqa: E221
+        cuda_common_gpu_architectures  = ['3.0', '3.5', '5.0']           # noqa: E221
+        cuda_hi_limit_gpu_architecture = None                            # noqa: E221
+        cuda_lo_limit_gpu_architecture = '2.0'                           # noqa: E221
+        cuda_all_gpu_architectures     = ['3.0', '3.2', '3.5', '5.0']    # noqa: E221
 
         if version_compare(cuda_version, '<7.0'):
-            cuda_limit_gpu_architecture = '5.2'
+            cuda_hi_limit_gpu_architecture = '5.2'
 
         if version_compare(cuda_version, '>=7.0'):
             cuda_known_gpu_architectures  += ['Kepler+Tegra', 'Kepler+Tesla', 'Maxwell+Tegra']  # noqa: E221
@@ -148,7 +177,7 @@ class CudaModule(ExtensionModule):
 
             if version_compare(cuda_version, '<8.0'):
                 cuda_common_gpu_architectures += ['5.2+PTX']  # noqa: E221
-                cuda_limit_gpu_architecture    = '6.0'        # noqa: E221
+                cuda_hi_limit_gpu_architecture = '6.0'        # noqa: E221
 
         if version_compare(cuda_version, '>=8.0'):
             cuda_known_gpu_architectures  += ['Pascal', 'Pascal+Tegra']  # noqa: E221
@@ -157,23 +186,45 @@ class CudaModule(ExtensionModule):
 
             if version_compare(cuda_version, '<9.0'):
                 cuda_common_gpu_architectures += ['6.1+PTX']  # noqa: E221
-                cuda_limit_gpu_architecture    = '7.0'        # noqa: E221
+                cuda_hi_limit_gpu_architecture = '7.0'        # noqa: E221
 
         if version_compare(cuda_version, '>=9.0'):
-            cuda_known_gpu_architectures  += ['Volta', 'Xavier']                   # noqa: E221
-            cuda_common_gpu_architectures += ['7.0', '7.0+PTX']                    # noqa: E221
-            cuda_all_gpu_architectures    += ['7.0', '7.0+PTX', '7.2', '7.2+PTX']  # noqa: E221
+            cuda_known_gpu_architectures  += ['Volta', 'Xavier'] # noqa: E221
+            cuda_common_gpu_architectures += ['7.0']             # noqa: E221
+            cuda_all_gpu_architectures    += ['7.0', '7.2']      # noqa: E221
+            # https://docs.nvidia.com/cuda/archive/9.0/cuda-toolkit-release-notes/index.html#unsupported-features
+            cuda_lo_limit_gpu_architecture = '3.0'               # noqa: E221
 
             if version_compare(cuda_version, '<10.0'):
-                cuda_limit_gpu_architecture = '7.5'
+                cuda_common_gpu_architectures += ['7.2+PTX']  # noqa: E221
+                cuda_hi_limit_gpu_architecture = '8.0'        # noqa: E221
 
         if version_compare(cuda_version, '>=10.0'):
-            cuda_known_gpu_architectures  += ['Turing']          # noqa: E221
-            cuda_common_gpu_architectures += ['7.5', '7.5+PTX']  # noqa: E221
-            cuda_all_gpu_architectures    += ['7.5', '7.5+PTX']  # noqa: E221
+            cuda_known_gpu_architectures  += ['Turing'] # noqa: E221
+            cuda_common_gpu_architectures += ['7.5']    # noqa: E221
+            cuda_all_gpu_architectures    += ['7.5']    # noqa: E221
 
             if version_compare(cuda_version, '<11.0'):
-                cuda_limit_gpu_architecture = '8.0'
+                cuda_common_gpu_architectures += ['7.5+PTX']  # noqa: E221
+                cuda_hi_limit_gpu_architecture = '8.0'        # noqa: E221
+
+        if version_compare(cuda_version, '>=11.0'):
+            cuda_known_gpu_architectures  += ['Ampere'] # noqa: E221
+            cuda_common_gpu_architectures += ['8.0']    # noqa: E221
+            cuda_all_gpu_architectures    += ['8.0']    # noqa: E221
+            # https://docs.nvidia.com/cuda/archive/11.0/cuda-toolkit-release-notes/index.html#deprecated-features
+            cuda_lo_limit_gpu_architecture = '3.5'      # noqa: E221
+
+            if version_compare(cuda_version, '<11.1'):
+                cuda_common_gpu_architectures += ['8.0+PTX']  # noqa: E221
+                cuda_hi_limit_gpu_architecture = '8.6'        # noqa: E221
+
+        if version_compare(cuda_version, '>=11.1'):
+            cuda_common_gpu_architectures += ['8.6', '8.6+PTX']  # noqa: E221
+            cuda_all_gpu_architectures    += ['8.6']             # noqa: E221
+
+            if version_compare(cuda_version, '<12.0'):
+                cuda_hi_limit_gpu_architecture = '9.0'        # noqa: E221
 
         if not cuda_arch_list:
             cuda_arch_list = 'Auto'
@@ -188,16 +239,10 @@ class CudaModule(ExtensionModule):
                     cuda_arch_list = detected
                 else:
                     cuda_arch_list = self._break_arch_string(detected)
-
-                if cuda_limit_gpu_architecture:
-                    filtered_cuda_arch_list = []
-                    for arch in cuda_arch_list:
-                        if arch:
-                            if version_compare(arch, '>=' + cuda_limit_gpu_architecture):
-                                arch = cuda_common_gpu_architectures[-1]
-                            if arch not in filtered_cuda_arch_list:
-                                filtered_cuda_arch_list.append(arch)
-                    cuda_arch_list = filtered_cuda_arch_list
+                cuda_arch_list = self._filter_cuda_arch_list(cuda_arch_list,
+                                                             cuda_lo_limit_gpu_architecture,
+                                                             cuda_hi_limit_gpu_architecture,
+                                                             cuda_common_gpu_architectures[-1])
             else:
                 cuda_arch_list = cuda_common_gpu_architectures
         elif isinstance(cuda_arch_list, str):
@@ -229,6 +274,7 @@ class CudaModule(ExtensionModule):
                     'Volta':         (['7.0'],             ['7.0']),
                     'Xavier':        (['7.2'],             []),
                     'Turing':        (['7.5'],             ['7.5']),
+                    'Ampere':        (['8.0'],             ['8.0']),
                 }.get(arch_name, (None, None))
 
             if arch_bin is None:
@@ -242,10 +288,6 @@ class CudaModule(ExtensionModule):
                     arch_ptx = arch_bin
                 cuda_arch_ptx += arch_ptx
 
-        cuda_arch_bin = re.sub('\\.', '', ' '.join(cuda_arch_bin))
-        cuda_arch_ptx = re.sub('\\.', '', ' '.join(cuda_arch_ptx))
-        cuda_arch_bin = re.findall('[0-9()]+', cuda_arch_bin)
-        cuda_arch_ptx = re.findall('[0-9]+',   cuda_arch_ptx)
         cuda_arch_bin = sorted(list(set(cuda_arch_bin)))
         cuda_arch_ptx = sorted(list(set(cuda_arch_ptx)))
 
@@ -253,15 +295,37 @@ class CudaModule(ExtensionModule):
         nvcc_archs_readable = []
 
         for arch in cuda_arch_bin:
-            m = re.match('([0-9]+)\\(([0-9]+)\\)', arch)
-            if m:
-                nvcc_flags += ['-gencode', 'arch=compute_' + m[2] + ',code=sm_' + m[1]]
-                nvcc_archs_readable += ['sm_' + m[1]]
+            arch, codev = re.fullmatch(
+                '([0-9]+\\.[0-9])(?:\\(([0-9]+\\.[0-9])\\))?', arch).groups()
+
+            if version_compare(arch, '<' + cuda_lo_limit_gpu_architecture):
+                continue
+            if version_compare(arch, '>=' + cuda_hi_limit_gpu_architecture):
+                continue
+
+            if codev:
+                arch = arch.replace('.', '')
+                codev = codev.replace('.', '')
+                nvcc_flags += ['-gencode', 'arch=compute_' + codev + ',code=sm_' + arch]
+                nvcc_archs_readable += ['sm_' + arch]
             else:
+                arch = arch.replace('.', '')
                 nvcc_flags += ['-gencode', 'arch=compute_' + arch + ',code=sm_' + arch]
                 nvcc_archs_readable += ['sm_' + arch]
 
         for arch in cuda_arch_ptx:
+            arch, codev = re.fullmatch(
+                '([0-9]+\\.[0-9])(?:\\(([0-9]+\\.[0-9])\\))?', arch).groups()
+
+            if codev:
+                arch = codev
+
+            if version_compare(arch, '<' + cuda_lo_limit_gpu_architecture):
+                continue
+            if version_compare(arch, '>=' + cuda_hi_limit_gpu_architecture):
+                continue
+
+            arch = arch.replace('.', '')
             nvcc_flags += ['-gencode', 'arch=compute_' + arch + ',code=compute_' + arch]
             nvcc_archs_readable += ['compute_' + arch]
 
diff --git a/test cases/cuda/3 cudamodule/meson.build b/test cases/cuda/3 cudamodule/meson.build
index f556328..fd5e83f 100644
--- a/test cases/cuda/3 cudamodule/meson.build
+++ b/test cases/cuda/3 cudamodule/meson.build
@@ -3,9 +3,9 @@ project('cudamodule', 'cuda', version : '1.0.0')
 nvcc = meson.get_compiler('cuda')
 cuda = import('unstable-cuda')
 
-arch_flags     = cuda.nvcc_arch_flags(nvcc, 'Auto', detected: ['6.0'])
-arch_readable  = cuda.nvcc_arch_readable(nvcc, 'Auto', detected: ['6.0'])
-driver_version = cuda.min_driver_version(nvcc)
+arch_flags     = cuda.nvcc_arch_flags(nvcc.version(), 'Auto', detected: ['6.0'])
+arch_readable  = cuda.nvcc_arch_readable(nvcc.version(), 'Auto', detected: ['6.0'])
+driver_version = cuda.min_driver_version(nvcc.version())
 
 message('NVCC version:   ' + nvcc.version())
 message('NVCC flags:     ' + ' '.join(arch_flags))
@@ -14,3 +14,57 @@ message('Driver version: >=' + driver_version)
 
 exe = executable('prog', 'prog.cu', cuda_args: arch_flags)
 test('cudatest', exe)
+
+
+#
+# Assert Series
+#
+
+# Sanity test.
+assert(' '.join(cuda.nvcc_arch_flags('11.1', '8.6')) ==
+       '-gencode arch=compute_86,code=sm_86')
+
+# CUDA Toolkit too old, flag filtered out.
+assert(' '.join(cuda.nvcc_arch_flags('11.0', '8.6')) ==
+       '')
+
+# Named architectures.
+assert(' '.join(cuda.nvcc_arch_flags('11.0', 'Ampere')) ==
+       '-gencode arch=compute_80,code=sm_80')
+
+# Splitting & deduplication.
+assert(' '.join(cuda.nvcc_arch_flags('11.0', 'Ampere;8.0,8.0')) ==
+       '-gencode arch=compute_80,code=sm_80')
+
+# Same, but list supplied as list.
+assert(' '.join(cuda.nvcc_arch_flags('11.0', ['Ampere', '8.0', '8.0'])) ==
+       '-gencode arch=compute_80,code=sm_80')
+
+# Same, but mode set to Auto with detected set to a string with a variety of separators.
+assert(' '.join(cuda.nvcc_arch_flags('11.0', 'Auto', detected: 'Ampere;8.0,8.0')) ==
+       '-gencode arch=compute_80,code=sm_80')
+
+# Same, but detected set to a list.
+assert(' '.join(cuda.nvcc_arch_flags('11.0', 'Auto', detected: ['Ampere', '8.0', '8.0'])) ==
+       '-gencode arch=compute_80,code=sm_80')
+
+# Ask for 8.6 binary with 8.0-level PTX.
+assert(' '.join(cuda.nvcc_arch_flags('11.1', '8.6(8.0)')) ==
+       '-gencode arch=compute_80,code=sm_86')
+
+# Same, but keep the 8.0 PTX.
+assert(' '.join(cuda.nvcc_arch_flags('11.1', '8.6(8.0)+PTX')) ==
+       '-gencode arch=compute_80,code=sm_86 -gencode arch=compute_80,code=compute_80')
+
+# Detected Ampere RTX 3090 on CUDA 10.2, saturate to 7.5+PTX
+assert(' '.join(cuda.nvcc_arch_flags('10.2', 'Auto', detected: ['8.0'])) ==
+       '-gencode arch=compute_75,code=sm_75 -gencode arch=compute_75,code=compute_75')
+
+# Failed to auto-detect with CUDA 10.2, default to common GPUs (3.0;3.5;5.0;5.2;6.0;6.1;7.0;7.5+PTX)
+assert(' '.join(cuda.nvcc_arch_flags('10.2', 'Auto', detected: [])) ==
+       '-gencode arch=compute_30,code=sm_30 -gencode arch=compute_35,code=sm_35 '+
+       '-gencode arch=compute_50,code=sm_50 -gencode arch=compute_52,code=sm_52 '+
+       '-gencode arch=compute_60,code=sm_60 -gencode arch=compute_61,code=sm_61 '+
+       '-gencode arch=compute_70,code=sm_70 -gencode arch=compute_75,code=sm_75 '+
+       '-gencode arch=compute_75,code=compute_75')
+