Test loop fusion implementation

test/test_loop_fusion.py

@@ -0,0 +1,276 @@
+__copyright__ = "Copyright (C) 2021 Kaushik Kulkarni"
+
+__license__ = """
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+"""
+
+import logging
+import sys
+
+import numpy as np
+
+import pyopencl as cl
+import pyopencl.clmath  # noqa
+import pyopencl.clrandom  # noqa
+
+import loopy as lp
+
+
+logger = logging.getLogger(__name__)
+
+try:
+    import faulthandler
+except ImportError:
+    pass
+else:
+    faulthandler.enable()
+
+from pyopencl.tools import pytest_generate_tests_for_pyopencl as pytest_generate_tests
+
+from loopy.version import LOOPY_USE_LANGUAGE_VERSION_2018_2  # noqa
+
+
+__all__ = [
+    "cl",  # "cl.create_some_context"
+    "pytest_generate_tests"
+]
+
+
+def test_loop_fusion_vanilla(ctx_factory):
+    ctx = ctx_factory()
+
+    knl = lp.make_kernel(
+        "{[i0, i1, j0, j1]: 0 <= i0, i1, j0, j1 < 10}",
+        """
+        a[i0]     = 1
+        b[i1, j0] = 2  {id=write_b}
+        c[j1]     = 3  {id=write_c}
+        """)
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                    frozenset(["j0", "j1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+    assert len(ref_knl["loopy_kernel"].all_inames()) == 4
+    assert len(knl["loopy_kernel"].all_inames()) == 3
+    assert len(knl["loopy_kernel"].id_to_insn["write_b"].within_inames
+               & knl["loopy_kernel"].id_to_insn["write_c"].within_inames) == 1
+
+    lp.auto_test_vs_ref(ref_knl, ctx, knl)
+
+
+def test_loop_fusion_outer_iname_preventing_fusion(ctx_factory):
+    ctx = ctx_factory()
+
+    knl = lp.make_kernel(
+        "{[i0, j0, j1]: 0 <= i0, j0, j1 < 10}",
+        """
+        a[i0]     = 1
+        b[i0, j0] = 2 {id=write_b}
+        c[j1]     = 3 {id=write_c}
+        """)
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                    frozenset(["j0", "j1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+
+    assert len(knl["loopy_kernel"].all_inames()) == 3
+    assert len(knl["loopy_kernel"].all_inames()) == 3
+    assert len(knl["loopy_kernel"].id_to_insn["write_b"].within_inames
+               & knl["loopy_kernel"].id_to_insn["write_c"].within_inames) == 0
+
+    lp.auto_test_vs_ref(ref_knl, ctx, knl)
+
+
+def test_loop_fusion_with_loop_independent_deps(ctx_factory):
+    ctx = ctx_factory()
+
+    knl = lp.make_kernel(
+        "{[j0, j1]: 0 <= j0, j1 < 10}",
+        """
+        a[j0]  = 1
+        b[j1] = 2 * a[j1]
+        """, seq_dependencies=True)
+
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                    frozenset(["j0", "j1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+
+    assert len(ref_knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].all_inames()) == 1
+
+    lp.auto_test_vs_ref(ref_knl, ctx, knl)
+
+
+def test_loop_fusion_constrained_by_outer_loop_deps(ctx_factory):
+    ctx = ctx_factory()
+
+    knl = lp.make_kernel(
+        "{[j0, j1]: 0 <= j0, j1 < 10}",
+        """
+        a[j0] = 1         {id=write_a}
+        b     = 2         {id=write_b}
+        c[j1] = 2 * a[j1] {id=write_c}
+        """, seq_dependencies=True)
+
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                    frozenset(["j0", "j1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+
+    assert len(ref_knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].id_to_insn["write_a"].within_inames
+               & knl["loopy_kernel"].id_to_insn["write_c"].within_inames) == 0
+
+    lp.auto_test_vs_ref(ref_knl, ctx, knl)
+
+
+def test_loop_fusion_with_loop_carried_deps1(ctx_factory):
+
+    ctx = ctx_factory()
+    knl = lp.make_kernel(
+        "{[i0, i1]: 1<=i0, i1<10}",
+        """
+        x[i0] = i0         {id=first_write}
+        x[i1-1] = i1 ** 2  {id=second_write}
+        """,
+        seq_dependencies=True)
+
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                               frozenset(["i0",
+                                                                          "i1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+
+    assert len(ref_knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].all_inames()) == 1
+    assert len(knl["loopy_kernel"].id_to_insn["first_write"].within_inames
+               & knl["loopy_kernel"].id_to_insn["second_write"].within_inames) == 1
+
+    lp.auto_test_vs_ref(ref_knl, ctx, knl)
+
+
+def test_loop_fusion_with_loop_carried_deps2(ctx_factory):
+    ctx = ctx_factory()
+    knl = lp.make_kernel(
+        "{[i0, i1]: 1<=i0, i1<10}",
+        """
+        x[i0-1] = i0     {id=first_write}
+        x[i1] = i1 ** 2  {id=second_write}
+        """,
+        seq_dependencies=True)
+
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                               frozenset(["i0",
+                                                                          "i1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+
+    assert len(ref_knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].id_to_insn["first_write"].within_inames
+               & knl["loopy_kernel"].id_to_insn["second_write"].within_inames) == 0
+
+    lp.auto_test_vs_ref(ref_knl, ctx, knl)
+
+
+def test_loop_fusion_with_indirection(ctx_factory):
+    ctx = ctx_factory()
+    map_ = np.random.permutation(10)
+    cq = cl.CommandQueue(ctx)
+
+    knl = lp.make_kernel(
+        "{[i0, i1]: 0<=i0, i1<10}",
+        """
+        x[i0] = i0            {id=first_write}
+        x[map[i1]] = i1 ** 2  {id=second_write}
+        """,
+        seq_dependencies=True)
+
+    ref_knl = knl
+
+    fused_chunks = lp.get_kennedy_unweighted_fusion_candidates(knl["loopy_kernel"],
+                                                               frozenset(["i0",
+                                                                          "i1"]))
+
+    knl = knl.with_kernel(lp.rename_inames_in_batch(knl["loopy_kernel"],
+                                                    fused_chunks))
+
+    assert len(ref_knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].all_inames()) == 2
+    assert len(knl["loopy_kernel"].id_to_insn["first_write"].within_inames
+               & knl["loopy_kernel"].id_to_insn["second_write"].within_inames) == 0
+
+    _, (out1,) = ref_knl(cq, map=map_)
+    _, (out2,) = knl(cq, map=map_)
+    np.testing.assert_allclose(out1, out2)
+
+
+def test_loop_fusion_with_induced_dependencies_from_sibling_nests(ctx_factory):
+    ctx = ctx_factory()
+    t_unit = lp.make_kernel(
+        "{[i0, j, i1, i2]: 0<=i0, j, i1, i2<10}",
+        """
+        <> tmp0[i0] = i0
+        <> tmp1[j] = tmp0[j]
+        <> tmp2[j] = j
+        out1[i1] = tmp2[i1]
+        out2[i2] = 2 * tmp1[i2]
+        """)
+    ref_t_unit = t_unit
+    knl = t_unit.default_entrypoint
+    knl = lp.rename_inames_in_batch(
+        knl,
+        lp.get_kennedy_unweighted_fusion_candidates(
+            knl, frozenset(["i0", "i1"])))
+    t_unit = t_unit.with_kernel(knl)
+
+    # 'i1', 'i2' should not be fused. If fused that would lead to an
+    # unshcedulable kernel. Making sure that the kernel 'runs' suffices that
+    # the transformation was successful.
+    lp.auto_test_vs_ref(ref_t_unit, ctx, t_unit)
+
+
+if __name__ == "__main__":
+    if len(sys.argv) > 1:
+        exec(sys.argv[1])
+    else:
+        from pytest import main
+        main([__file__])
+
+# vim: fdm=marker