legate.sparse tracking branch [DO NOT MERGE]

conda/environment-cluster-3.10.yml

@@ -0,0 +1,57 @@
+name: cunumeric-test
+channels:
+  - conda-forge
+dependencies:
+  - python=3.10
+
+  # build
+  - git
+  - nccl
+  - make
+  - zlib
+  - cmake==3.24
+  - ninja
+  - sysroot_linux-64==2.17 # [linux64]
+  - setuptools>=60
+  - cutensor>=1.3.3
+  - scikit-build>=0.13.1
+
+  # runtime
+  - cffi
+  - numpy>=1.22
+  - opt_einsum
+  - pyarrow>=5
+  - scipy
+  - typing_extensions
+  - llvm-openmp
+  - openblas=*=*openmp*
+
+  # tests
+  - clang>=8
+  - clang-tools>=8
+  - colorama
+  - coverage
+  - mock
+  - mypy>=0.961
+  - pre-commit
+  - pynvml
+  - pytest
+  - pytest-cov
+  - pytest-mock
+  - pytest-lazy-fixture
+  - types-docutils
+
+  # pip dependencies
+  - pip
+  - pip:
+    # docs
+    - jinja2
+    - pydata-sphinx-theme
+    - recommonmark
+    - markdown<3.4.0
+    - sphinx>=4.4.0
+    - sphinx-copybutton
+    - sphinx-markdown-tables
+
+    # examples
+    - tifffile

conda/environment-cluster-3.8.yml

@@ -0,0 +1,57 @@
+name: cunumeric-test
+channels:
+  - conda-forge
+dependencies:
+  - python=3.8
+
+  # build
+  - git
+  - nccl
+  - make
+  - zlib
+  - cmake==3.24
+  - ninja
+  - sysroot_linux-64==2.17 # [linux64]
+  - setuptools>=60
+  - cutensor>=1.3.3
+  - scikit-build>=0.13.1
+
+  # runtime
+  - cffi
+  - numpy>=1.22
+  - opt_einsum
+  - pyarrow>=5
+  - scipy
+  - typing_extensions
+  - llvm-openmp
+  - openblas=*=*openmp*
+
+  # tests
+  - clang>=8
+  - clang-tools>=8
+  - colorama
+  - coverage
+  - mock
+  - mypy>=0.961
+  - pre-commit
+  - pynvml
+  - pytest
+  - pytest-cov
+  - pytest-mock
+  - pytest-lazy-fixture
+  - types-docutils
+
+  # pip dependencies
+  - pip
+  - pip:
+    # docs
+    - jinja2
+    - pydata-sphinx-theme
+    - recommonmark
+    - markdown<3.4.0
+    - sphinx>=4.4.0
+    - sphinx-copybutton
+    - sphinx-markdown-tables
+
+    # examples
+    - tifffile

conda/environment-cluster-3.9.yml

@@ -0,0 +1,57 @@
+name: cunumeric-test
+channels:
+  - conda-forge
+dependencies:
+  - python=3.9
+
+  # build
+  - git
+  - nccl
+  - make
+  - zlib
+  - cmake==3.24
+  - ninja
+  - sysroot_linux-64==2.17 # [linux64]
+  - setuptools>=60
+  - cutensor>=1.3.3
+  - scikit-build>=0.13.1
+
+  # runtime
+  - cffi
+  - numpy>=1.22
+  - opt_einsum
+  - pyarrow>=5
+  - scipy
+  - typing_extensions
+  - llvm-openmp
+  - openblas=*=*openmp*
+
+  # tests
+  - clang>=8
+  - clang-tools>=8
+  - colorama
+  - coverage
+  - mock
+  - mypy>=0.961
+  - pre-commit
+  - pynvml
+  - pytest
+  - pytest-cov
+  - pytest-mock
+  - pytest-lazy-fixture
+  - types-docutils
+
+  # pip dependencies
+  - pip
+  - pip:
+    # docs
+    - jinja2
+    - pydata-sphinx-theme
+    - recommonmark
+    - markdown<3.4.0
+    - sphinx>=4.4.0
+    - sphinx-copybutton
+    - sphinx-markdown-tables
+
+    # examples
+    - tifffile

cunumeric/module.py

@@ -1130,6 +1130,21 @@ def reshape(
     return a.reshape(newshape, order=order)
 
 
+# Support a very basic (and incorrect in the general case) version of
+# broadcast_to that only works if we don't perform direct writes to
+# the underlying array.
+@add_boilerplate("array")
+def broadcast_to(
+    array: ndarray, shape: NdShapeLike, subok: bool = False
+) -> ndarray:
+    from .deferred import DeferredArray
+
+    thunk = array._thunk
+    store = thunk._broadcast(shape)  # type: ignore
+    new_thunk = DeferredArray(thunk.runtime, store, array.dtype)
+    return ndarray(shape, dtype=array.dtype, thunk=new_thunk)
+
+
 # Transpose-like operations
 
 

cunumeric/runtime.py

@@ -462,17 +462,37 @@ def find_or_create_array_thunk(
                 shape=array.shape,
                 optimize_scalar=False,
             )
-            store.attach_external_allocation(
-                self.legate_context,
-                array.data,
-                share,
-            )
-            return DeferredArray(
-                self,
-                store,
-                dtype=array.dtype,
-                numpy_array=array if share else None,
-            )
+            # If the attach hack is enabled, do writes from the input array
+            # into the resulting array. Otherwise, do an attach. The attach
+            # hack enables scaling to multiple nodes when restricted
+            # attaches must be performed.
+            if settings.attach_hack():
+                assert not share
+                thunk = DeferredArray(
+                    self,
+                    store,
+                    dtype=array.dtype,
+                )
+                for index in np.ndindex(array.shape):
+                    thunk.set_item(
+                        index,
+                        self.create_wrapped_scalar(
+                            array[index], array.dtype, ()
+                        ),
+                    )
+                return thunk
+            else:
+                store.attach_external_allocation(
+                    self.legate_context,
+                    array.data,
+                    share,
+                )
+                return DeferredArray(
+                    self,
+                    store,
+                    dtype=array.dtype,
+                    numpy_array=array if share else None,
+                )
 
         assert not defer
         # Make this into an eager evaluated thunk

cunumeric/settings.py

@@ -53,6 +53,18 @@ class CunumericRuntimeSettings(Settings):
         """,
     )
 
+    attach_hack: PrioritizedSetting[bool] = PrioritizedSetting(
+        "attach-hack",
+        "CUNUMERIC_ATTACH_HACK",
+        default=False,
+        convert=convert_bool,
+        help="""
+        Enable the attach hack which avoids attaching in certain cases by doing
+        direct writes. This may be needed for multi-node runs of some
+        applications.
+        """,
+    )
+
     report_coverage: PrioritizedSetting[bool] = PrioritizedSetting(
         "report_coverage",
         "CUNUMERIC_REPORT_COVERAGE",