Skip to content

Commit 95ba959

Browse files
Copilotggorman
Copilot
and
ggorman
committed
Add test and documentation for linear clause fix
Co-authored-by: ggorman <5394691+ggorman@users.noreply.github.com>
1 parent fcc8544 commit 95ba959

5 files changed

Lines changed: 309 additions & 436 deletions

File tree

‎scripts/demonstrate_fix.py‎

Lines changed: 174 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,174 @@
1+
#!/usr/bin/env python3
2+
"""
3+
Demonstration of the OpenMP linear clause fix for issue #320.
4+
5+
This script shows how the fix addresses the GCC compilation issue
6+
with OpenMP SIMD pragmas in non-canonical loop forms.
7+
"""
8+
9+
def demonstrate_fix():
10+
"""
11+
Show the key components of the fix and what it accomplishes.
12+
"""
13+
print("OpenMP Linear Clause Fix for GCC Compatibility (Issue #320)")
14+
print("=" * 60)
15+
16+
print("\n1. PROBLEM DESCRIPTION:")
17+
print(" - GCC doesn't accept 'pragma omp simd' with non-canonical loop forms")
18+
print(" - This occurs when blockinner=True creates complex nested loops")
19+
print(" - ICC is more permissive and doesn't have this issue")
20+
print(" - Needed for GCC >= 4.9 (which supports OpenMP 4.0)")
21+
22+
print("\n2. SOLUTION IMPLEMENTED:")
23+
print(" - Added OpenMP 'linear' clause to SIMD pragmas when needed")
24+
print(" - Linear clause tells OpenMP that variables change linearly with iteration")
25+
print(" - Only applied for GCC compiler, not ICC")
26+
print(" - Automatically detects when linear clauses are needed")
27+
28+
print("\n3. PRAGMA TRANSFORMATIONS:")
29+
print(" Before (problematic for GCC):")
30+
print(" #pragma omp simd")
31+
print(" for (int i = ...)")
32+
print(" for (int j = ...)")
33+
print()
34+
print(" After (GCC compatible):")
35+
print(" #pragma omp simd linear(i,j)")
36+
print(" for (int i = ...)")
37+
print(" for (int j = ...)")
38+
39+
print("\n4. NEW PRAGMA VARIANTS ADDED:")
40+
variants = [
41+
("simd-for", "Basic SIMD pragma"),
42+
("simd-for-linear", "SIMD with linear clause"),
43+
("simd-for-aligned-linear", "SIMD with both aligned and linear clauses")
44+
]
45+
46+
for variant, description in variants:
47+
print(f" - {variant}: {description}")
48+
49+
print("\n5. DETECTION LOGIC:")
50+
conditions = [
51+
"Using GCC compiler (not ICC)",
52+
"GCC version >= 4.9 (OpenMP 4.0 support)",
53+
"Complex nested loop structure detected (3+ levels)",
54+
"Block dimensions or common loop indices present"
55+
]
56+
57+
for i, condition in enumerate(conditions, 1):
58+
print(f" {i}. {condition}")
59+
60+
print("\n6. CODE LOCATIONS MODIFIED:")
61+
files = [
62+
"devito/passes/iet/languages/openmp.py - Added new pragma variants",
63+
"devito/passes/iet/parpragma.py - Enhanced SIMD pragma generation logic"
64+
]
65+
66+
for file_info in files:
67+
print(f" - {file_info}")
68+
69+
print("\n7. BENEFITS:")
70+
benefits = [
71+
"GCC compilation works with blockinner=True and aggressive DSE",
72+
"ICC compatibility maintained (no linear clauses added unnecessarily)",
73+
"Automatic detection - no user configuration required",
74+
"Backward compatible with existing code"
75+
]
76+
77+
for benefit in benefits:
78+
print(f" ✓ {benefit}")
79+
80+
81+
def show_example_scenarios():
82+
"""
83+
Show example scenarios where the fix would be applied.
84+
"""
85+
print("\n" + "=" * 60)
86+
print("EXAMPLE SCENARIOS")
87+
print("=" * 60)
88+
89+
scenarios = [
90+
{
91+
"name": "TTI (Tilted Transverse Isotropy) with 3D blocking",
92+
"description": "Complex seismic stencil with blockinner=True",
93+
"triggers_fix": True,
94+
"reason": "Multiple nested loops with block dimensions"
95+
},
96+
{
97+
"name": "Simple 2D heat equation without blocking",
98+
"description": "Basic stencil with no blocking transformations",
99+
"triggers_fix": False,
100+
"reason": "No complex loop nesting"
101+
},
102+
{
103+
"name": "3D acoustic stencil with blockinner=True",
104+
"description": "3D space-order stencil with aggressive blocking",
105+
"triggers_fix": True,
106+
"reason": "Deep loop nesting from blocking"
107+
},
108+
{
109+
"name": "Same code compiled with ICC",
110+
"description": "Any stencil using Intel compiler",
111+
"triggers_fix": False,
112+
"reason": "ICC doesn't need linear clauses"
113+
}
114+
]
115+
116+
for i, scenario in enumerate(scenarios, 1):
117+
print(f"\n{i}. {scenario['name']}")
118+
print(f" Description: {scenario['description']}")
119+
print(f" Triggers fix: {'Yes' if scenario['triggers_fix'] else 'No'}")
120+
print(f" Reason: {scenario['reason']}")
121+
122+
123+
def show_testing_approach():
124+
"""
125+
Explain how the fix can be tested.
126+
"""
127+
print("\n" + "=" * 60)
128+
print("TESTING APPROACH")
129+
print("=" * 60)
130+
131+
tests = [
132+
{
133+
"type": "Unit Tests",
134+
"description": "Test pragma generation logic in isolation",
135+
"file": "test_linear_clause_unit.py"
136+
},
137+
{
138+
"type": "Integration Tests",
139+
"description": "Test with actual Devito operators and blocking",
140+
"file": "test_linear_pragma_fix.py"
141+
},
142+
{
143+
"type": "Regression Tests",
144+
"description": "Ensure existing functionality still works",
145+
"file": "Existing test suite (test_dle.py, etc.)"
146+
},
147+
{
148+
"type": "Compiler Tests",
149+
"description": "Test with both GCC and ICC to ensure compatibility",
150+
"file": "Manual testing with different compilers"
151+
}
152+
]
153+
154+
for test in tests:
155+
print(f"\n• {test['type']}")
156+
print(f" {test['description']}")
157+
print(f" Location: {test['file']}")
158+
159+
160+
if __name__ == "__main__":
161+
demonstrate_fix()
162+
show_example_scenarios()
163+
show_testing_approach()
164+
165+
print("\n" + "=" * 60)
166+
print("SUMMARY")
167+
print("=" * 60)
168+
print("This fix resolves issue #320 by automatically adding OpenMP 'linear'")
169+
print("clauses to SIMD pragmas when using GCC with complex loop structures")
170+
print("from blocking transformations. The fix is conservative, automatic,")
171+
print("and maintains compatibility with both GCC and ICC compilers.")
172+
print()
173+
print("The implementation is ready for testing with TTI examples and")
174+
print("other complex stencils that use blockinner=True with aggressive DSE.")

‎test_blockinner_issue.py‎

Lines changed: 0 additions & 135 deletions
This file was deleted.

0 commit comments

Comments
 (0)