CAKE2 Tests

Test suite for CAKE2 local mode functionality.

Prerequisites

• Local CANN environment setup
• NPU hardware available
• Python 3.10+
• torch_npu installed: pip install torch_npu

Running Tests

Test Local Compilation

This test verifies that CAKE2 can generate and compile AscendC operators locally.

# Run local compilation test
python3 tests/test_local_compilation.py

GELUCustom Test Case

The main test case validates the complete CAKE2 pipeline using a GELUCustom operator:

1. Environment Check

   • Verify CANN installation
   • Check NPU device availability
   • Validate torch_npu setup

2. Project Generation

   • Generate AscendC project structure
   • Create CMakeLists.txt and build scripts
   • Generate kernel and host code

3. Compilation

   • Build project with cmake
   • Link against CANN libraries
   • Generate custom operator library

4. Evaluation

   • Import compiled operator
   • Run correctness tests vs PyTorch
   • Measure performance speedup

Expected Output

============================================================
CAKE2 Local Mode Test Suite
============================================================
Testing: GELUCustom operator
============================================================

============================================================
Test: Environment Check
============================================================

→ Checking CANN installation...
✓ ASCEND_HOME_PATH: /usr/local/Ascend/ascend-toolkit/latest
✓ CANN environment variables set

→ Checking NPU devices...
✓ Found 8 NPU devices
✓ Using device: 0

============================================================
✅ Test PASSED: Environment check
============================================================

============================================================
Test: Project Generation
============================================================

✓ Created test operator: gelu_custom
✓ Generated project structure:
  - CMakeLists.txt
  - build.sh
  - op_host/
  - op_kernel/

→ Verifying project files...
✓ Found: GeluCustom/CMakeLists.txt
✓ Found: GeluCustom/op_host/gelu_custom.cpp
✓ Found: GeluCustom/op_kernel/gelu_custom.cpp

============================================================
✅ Test PASSED: Project generation
============================================================

============================================================
Test: Local Compilation
============================================================

→ Building project with cmake...
✓ CMake configuration successful
✓ Compilation successful
✓ Generated library: libgelu_custom.so

============================================================
✅ Test PASSED: Compilation
============================================================

============================================================
Test: Evaluation
============================================================

→ Testing correctness...
✓ Import successful
✓ Correctness test: 100.00% match
✓ Max difference: 1.2e-06

→ Testing performance...
✓ PyTorch time: 0.234ms
✓ Custom time: 0.087ms
✓ Speedup: 2.69x

============================================================
✅ Test PASSED: Evaluation
============================================================

============================================================
Test Summary
============================================================
✅ PASS: Environment Check
✅ PASS: Project Generation
✅ PASS: Compilation
✅ PASS: Evaluation
============================================================
Results: 4/4 tests passed
============================================================

Test Coverage

These tests ensure that:

1. CANN Environment: Proper CANN setup and environment variables
2. NPU Hardware: NPU devices are available and accessible
3. Code Generation: AscendC projects are generated correctly
4. Compilation: Projects compile successfully with CANN
5. Correctness: Generated operators produce correct results
6. Performance: Operators achieve expected speedup

Troubleshooting

CANN Not Found

✗ Environment check failed: ASCEND_HOME_PATH not set

Solution: Set up CANN environment:

export ASCEND_HOME_PATH=/usr/local/Ascend/ascend-toolkit/latest
source /usr/local/Ascend/ascend-toolkit/set_env.sh

NPU Not Available

✗ No NPU devices found

Solution: Check NPU status:

npu-smi info
# If devices are busy, reset them:
npu-smi -i 0 -r

Compilation Failed

✗ Compilation failed: CMake error

Solution: Check build environment:

# Verify CANN libraries
ls $ASCEND_HOME_PATH/lib
# Check cmake version
cmake --version  # Should be >= 3.16

Import Failed

✗ Import failed: cannot open shared object file

Solution: Check library path:

# Add to LD_LIBRARY_PATH
export LD_LIBRARY_PATH=$ASCEND_HOME_PATH/lib:$LD_LIBRARY_PATH
# Verify library exists
ls build/libgelu_custom.so

GELUCustom Implementation Details

The test uses a GELU (Gaussian Error Linear Unit) activation function:

Formula: GELU(x) = 0.5 * x * (1 + tanh(sqrt(2/π) * (x + 0.044715 * x³)))

Test Parameters:

• Input shape: [1024, 4096]
• Data type: float32
• Target accuracy: > 99.99%
• Target speedup: > 1.5x

Key Features Tested:

• Element-wise computation
• Memory tiling strategy
• SIMD vectorization
• Numerical stability

Adding New Test Cases

To add new operator tests, extend test_local_compilation.py:

async def test_new_operator():
    """Test new operator implementation"""
    print("=" * 60)
    print("Test: New Operator")
    print("=" * 60)

    # Operator-specific test logic
    # 1. Generate project
    # 2. Compile
    # 3. Evaluate
    # 4. Verify results

    return True  # or False if test fails

Performance Benchmarks

Expected performance ranges for different operator types:

Operator Type	Expected Speedup
Element-wise (GELU, ReLU)	2-5x
Reductions (Sum, Mean)	1.5-3x
Matrix ops (MatMul)	1.2-2x
Fused ops	3-10x

CI/CD Integration

Local tests can be integrated into CI/CD pipelines:

# Example GitLab CI
test_cake2:
  script:
    - source /usr/local/Ascend/ascend-toolkit/set_env.sh
    - python3 tests/test_local_compilation.py
  only:
    - master
    - merge_requests