feat[next]: SDFGConvertible Program for dace_fieldview backend #1742

DropD · 2024-11-19T11:00:16Z

Description

Add a decrator.Program subclass, which implements SDFGConvertible to dace_fieldview backend, analogous to the one in dace_iterator. Conditionally shadow decorator.Program with it and reactivate the orchestration tests by using dace_fieldview instead of dace_iterator.

One caveat: The toolchain is not ready for pure CompileTimeConnectivities in all cases yet, so the test_sdfgConvertible_connectivities had to be adjusted for the moment.

Requirements

All fixes and/or new features come with corresponding tests.
Important design decisions have been documented in the approriate ADR inside the docs/development/ADRs/ folder.

If this PR contains code authored by new contributors please make sure:

The PR contains an updated version of the AUTHORS.md file adding the names of all the new contributors.

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

src/gt4py/next/ffront/decorator.py

edopao · 2024-11-19T11:36:46Z

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

+                if conn_id not in self.connectivity_tables_data_descriptors:
+                    conn = self.connectivities[name]
+                    self.connectivity_tables_data_descriptors[conn_id] = dace.data.Array(
+                        dtype=dace.int64 if conn.index_type == np.int64 else dace.int32,


I think that dace dtypes has a utility to parse numpy types:
dace.dtypes.typeclass(conn.index_type)

Trying that lead to KeyError: dtype('int64') in typeclass().

and trying dace.dtypes.typeclass(conn.index_type.type) leads to wrong stencil results. Edit: this might have instead been caused by auto_optimize=False.

edit: after merging main, it passes. However according to type hints of NeighborTableOffsetProvider, your version should work and mine should fail. Yet, in the tests, the connectivities are set up in a way so that my version works and yours fails. The current version on the other hand is safe in both cases.

If I can get the test to produce the correct types so the official version works, I will change it. Otherwise I would prefer to leave this one until the neighbor table construction is anyway changed by what @havogt is working on.

Could not get it to work, it seems like the type hints are wrong, but there is no point investing time in this when the API will change heavily anyway.

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

src/gt4py/next/program_processors/runners/dace_iterator/__init__.py

kotsaloscv · 2024-11-20T13:12:55Z

src/gt4py/next/program_processors/runners/dace_iterator/__init__.py

-                    continue
-                if param.id not in sdfg.gt4py_program_input_fields:
-                    continue
-                sdfg.offset_providers_per_input_field.setdefault(param.id, []).extend(list(shifts))


Same here.

All these properties are needed for the automatic halo exchange placement.

Results from discussion in a separate channel for "offset_providers_per_input_field":

it is correct, as only used in icon4py which uses "unstructured" fields. There is max one horizontal dimension per field.

trace_shifts should work™️ partially on GTIR (but not accross as_fieldop).

currently not actively used in icon4py, so defer implementation and put a TODO instead.

src/gt4py/next/program_processors/runners/dace_iterator/__init__.py

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

edopao · 2024-11-25T09:03:43Z

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

+        # Add them as dynamic properties to the SDFG
+        program = typing.cast(
+            itir.Program, gtir_stage.data
+        )  # we already checked that our backend uses GTIR


we already checked that our backend uses GTIR

where? otherwise we could make an assert isinstance().

We checked that self.backend is a DaCe backend. Probably should check that it is specifically the dace_fieldview backend.

I see, thank you for the explanation. No, not needed to check that it is the fieldview backend because the iterator backend will be removed soon (next week).

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

src/gt4py/next/program_processors/runners/dace_iterator/__init__.py

DropD · 2024-11-26T13:15:23Z

src/gt4py/next/ffront/decorator.py

+    connectivities: Optional[common.OffsetProvider] = (
+        None  # TODO(ricoh): replace with common.OffsetProviderType once the temporary pass doesn't require the runtime information
+    )


@havogt I had to change this to satisfy mypy. Now it is also consistent with CompileTimeArgs

makes sense

havogt · 2024-11-26T13:22:47Z

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

+        return {name for name in cls().visit(program) if name in field_param_names}
+
+
+class InputNamesExtractor(SymbolNameSetExtractor):


Could you move these utilities to some module in iterator/transforms. I would like to use them in other places.

I was wondering about that. Will move the classes as well as the tests.

The dace orchestration tests are temporarily skipped until #1742 is merged. The dace backend with SDFG optimization is temporarily disabled in unit tests until #1639 is merged. A second PR will reorganize the files in dace backend module.

edopao · 2024-12-04T21:06:13Z

If it can help, I got the tests to pass locally with this diff:

diff --git a/src/gt4py/next/program_processors/runners/dace_fieldview/program.py b/src/gt4py/next/program_processors/runners/dace_fieldview/program.py
index 803ae866..d65f4fac 100644
--- a/src/gt4py/next/program_processors/runners/dace_fieldview/program.py
+++ b/src/gt4py/next/program_processors/runners/dace_fieldview/program.py
@@ -176,7 +176,7 @@ class Program(decorator.Program, dace.frontend.python.common.SDFGConvertible):
                     conn = self.connectivities[name]
                     assert common.is_neighbor_table(conn)
                     self.connectivity_tables_data_descriptors[conn_id] = dace.data.Array(
-                        dtype=dace.int64 if conn.dtype == np.int64 else dace.int32,
+                        dtype=dace.dtypes.typeclass(conn.dtype.scalar_type),
                         shape=[
                             symbols[dace_utils.field_size_symbol_name(conn_id, 0)],
                             symbols[dace_utils.field_size_symbol_name(conn_id, 1)],
diff --git a/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py b/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py
index 9bde9aca..80cd9eac 100644
--- a/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py
+++ b/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py
@@ -133,7 +133,7 @@ def test_sdfgConvertible_connectivities(unstructured_case):
         data=xp.asarray([[0, 1], [1, 2], [2, 0]]),
         allocator=allocator,
     )
-    connectivities = {"E2V": e2v.__gt_type__()}
+    connectivities = {"E2V": e2v}
     offset_provider = OffsetProvider_t.dtype._typeclass.as_ctypes()(E2V=e2v.data_ptr())
 
     SDFG = sdfg.to_sdfg(connectivities=connectivities)

DropD · 2024-12-17T10:43:25Z

If it can help, I got the tests to pass locally with this diff:

diff --git a/src/gt4py/next/program_processors/runners/dace_fieldview/program.py b/src/gt4py/next/program_processors/runners/dace_fieldview/program.py
index 803ae866..d65f4fac 100644
--- a/src/gt4py/next/program_processors/runners/dace_fieldview/program.py
+++ b/src/gt4py/next/program_processors/runners/dace_fieldview/program.py
@@ -176,7 +176,7 @@ class Program(decorator.Program, dace.frontend.python.common.SDFGConvertible):
                     conn = self.connectivities[name]
                     assert common.is_neighbor_table(conn)
                     self.connectivity_tables_data_descriptors[conn_id] = dace.data.Array(
-                        dtype=dace.int64 if conn.dtype == np.int64 else dace.int32,
+                        dtype=dace.dtypes.typeclass(conn.dtype.scalar_type),
                         shape=[
                             symbols[dace_utils.field_size_symbol_name(conn_id, 0)],
                             symbols[dace_utils.field_size_symbol_name(conn_id, 1)],
diff --git a/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py b/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py
index 9bde9aca..80cd9eac 100644
--- a/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py
+++ b/tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py
@@ -133,7 +133,7 @@ def test_sdfgConvertible_connectivities(unstructured_case):
         data=xp.asarray([[0, 1], [1, 2], [2, 0]]),
         allocator=allocator,
     )
-    connectivities = {"E2V": e2v.__gt_type__()}
+    connectivities = {"E2V": e2v}
     offset_provider = OffsetProvider_t.dtype._typeclass.as_ctypes()(E2V=e2v.data_ptr())
 
     SDFG = sdfg.to_sdfg(connectivities=connectivities)

This is part of what I ended up doing but I wanted to honor the original purpose of the test, which was to compile the SDFG without knowledge of the connectivity tables and then call it with two different ones.

…or_for)

edopao

Some suggestions for cleanup, otherwise it looks good to me.

src/gt4py/next/iterator/transforms/extractors.py

tests/next_tests/integration_tests/feature_tests/dace/test_orchestration.py

…dfg-convertible

src/gt4py/next/program_processors/runners/dace_fieldview/program.py

kotsaloscv

LGTM, although we need to additionally check it with ICON4Py.

DropD added 2 commits November 18, 2024 13:38

[wip] start adding sdfg convertible Program for dace-fieldview

4c41f02

add SDFGConvertible Program replacement to dace_fieldview

481af34

edopao reviewed Nov 19, 2024

View reviewed changes

DropD added 4 commits November 19, 2024 15:09

improve generate_sdfg args, remove old Program replacement

ad0a7b2

Merge branch 'main' into gtir-sdfg-convertible

307b537

turn auto_optimize back on in __sdfg__

7c4d197

disable auto_opt once again in __sdfg__

3c56151

edopao reviewed Nov 20, 2024

View reviewed changes

src/gt4py/next/program_processors/runners/dace_fieldview/program.py Outdated Show resolved Hide resolved

support only CUDA device type in dace_fieldview Program

c5b4c43

kotsaloscv self-requested a review November 20, 2024 09:34

kotsaloscv requested changes Nov 20, 2024

View reviewed changes

edopao reviewed Nov 20, 2024

View reviewed changes

src/gt4py/next/program_processors/runners/dace_fieldview/program.py Outdated Show resolved Hide resolved

[wip] bring back extra sdfg attributes for halo placement

835c115

edopao reviewed Nov 25, 2024

View reviewed changes

DropD commented Nov 25, 2024

View reviewed changes

src/gt4py/next/program_processors/runners/dace_iterator/__init__.py Outdated Show resolved Hide resolved

DropD added 4 commits November 26, 2024 11:16

partially add halo exchange helper attrs with tests

c126cb2

add dace/gt4py type parsing crosscheck

0b90583

Merge branch 'main' into gtir-sdfg-convertible

8c1f9f4

fix dace_fieldview.program tests

77c250b

DropD commented Nov 26, 2024

View reviewed changes

havogt reviewed Nov 26, 2024

View reviewed changes

edopao mentioned this pull request Nov 28, 2024

feat[next]: Remove dace_iterator backend and pass_manager_legacy #1753

Merged

DropD added 2 commits December 3, 2024 14:25

refactor extractors and remove debuginfo warning

f8ec8f5

Merge branch 'main' into gtir-sdfg-convertible

5831591

work around gtir transforms requiring connectivity tables

19c76ad

edopao added 2 commits December 18, 2024 11:59

Merge remote-tracking branch 'origin/main' into gtir-sdfg-convertible

b2931a5

Add visitor fir Literal node

fd7f472

edopao added 2 commits December 18, 2024 14:48

Fix for attribute rename itir -> gtir on latest main

71e54fd

Fix error on main (is_field_allocator_factory_for -> is_field_allocat…

99030d0

…or_for)

edopao reviewed Dec 19, 2024

View reviewed changes

DropD added 4 commits January 6, 2025 09:57

cover addition in the program tests

439ec43

Merge branch 'gtir-sdfg-convertible-node-visitor-regtest' into gtir-s…

7853745

…dfg-convertible

clean up dace-program tests and extractors

6c6a4e3

cleanup orchestration tests

f799bc0

edopao reviewed Jan 7, 2025

View reviewed changes

src/gt4py/next/program_processors/runners/dace_fieldview/program.py Outdated Show resolved Hide resolved

kotsaloscv approved these changes Jan 8, 2025

View reviewed changes

edopao mentioned this pull request Jan 10, 2025

Test GTIR-DaCe backend C2SM/icon4py#638

Draft

edopao added 4 commits January 13, 2025 12:14

Fix attribute dtype error

9c884e3

fix mypy error

3ecaab4

fix stride problem

8976caa

fix stride problem (1)

d89383d

edopao merged commit 9a56fbd into GridTools:main Jan 13, 2025
21 checks passed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat[next]: SDFGConvertible Program for dace_fieldview backend #1742

feat[next]: SDFGConvertible Program for dace_fieldview backend #1742

DropD commented Nov 19, 2024 •

edited

Loading

edopao Nov 19, 2024

DropD Nov 19, 2024 •

edited

Loading

DropD Nov 19, 2024 •

edited

Loading

DropD Nov 19, 2024

DropD Nov 19, 2024

kotsaloscv Nov 20, 2024

DropD Nov 25, 2024

edopao Nov 25, 2024

DropD Nov 26, 2024

edopao Nov 26, 2024

DropD Nov 26, 2024

havogt Nov 26, 2024

havogt Nov 26, 2024

DropD Dec 3, 2024

edopao commented Dec 4, 2024

DropD commented Dec 17, 2024

edopao left a comment

kotsaloscv left a comment

		return {name for name in cls().visit(program) if name in field_param_names}


		class InputNamesExtractor(SymbolNameSetExtractor):

feat[next]: SDFGConvertible Program for dace_fieldview backend #1742

feat[next]: SDFGConvertible Program for dace_fieldview backend #1742

Conversation

DropD commented Nov 19, 2024 • edited Loading

Description

Requirements

Choose a reason for hiding this comment

DropD Nov 19, 2024 • edited Loading

Choose a reason for hiding this comment

DropD Nov 19, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

edopao commented Dec 4, 2024

DropD commented Dec 17, 2024

edopao left a comment

Choose a reason for hiding this comment

kotsaloscv left a comment

Choose a reason for hiding this comment

DropD commented Nov 19, 2024 •

edited

Loading

DropD Nov 19, 2024 •

edited

Loading

DropD Nov 19, 2024 •

edited

Loading