research/problems/gemm_optimization/near_tile/readme

GEMM Optimization Problem
=========================

Problem Setting
---------------
Design and optimize high-performance Triton kernels for General Matrix-Matrix Multiplication (GEMM) on GPU. This problem focuses on implementing efficient matrix multiplication kernels using Triton's JIT compilation system.

The challenge involves optimizing:
- **Memory access patterns**: Efficient loading and storing of matrix data
- **Block tiling**: Optimal block sizes for GPU execution
- **Autotuning**: Leveraging Triton's autotuning capabilities
- **Activation functions**: Implementing GELU activation within the kernel
- **Performance benchmarking**: Achieving speedup over baseline implementations

Target
------
- **Primary**: Maximize geometric mean speedup over baseline (higher is better)
- **Secondary**: Ensure correctness across diverse matrix shapes
- **Tertiary**: Minimize kernel launch overhead and memory usage

API Specification
-----------------
Implement a `Solution` class that returns a Triton kernel implementation:

```python
class Solution:
    def solve(self, spec_path: str = None) -> dict:
        """
        Returns a dict with either:
        - {"code": "python_code_string"}
        - {"program_path": "path/to/kernel.py"}
        """
        # Your implementation
        pass
```

Your kernel implementation must provide:

```python
import torch
import triton
import triton.language as tl

def matmul(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
    """
    Matrix multiplication with GELU activation.
    
    Args:
        a: Input tensor of shape (M, K)
        b: Input tensor of shape (K, N)
    
    Returns:
        Output tensor of shape (M, N) with GELU activation applied
    """
    pass
```

Required GELU Implementation:
```python
@triton.jit
def gelu(x):
    return x * 0.5 * (1.0 + tl.extra.cuda.libdevice.erf(x * 0.7071067811865476))
```

API Usage Notes
---------------
- The evaluator looks for a `matmul` function in the module namespace
- Function must handle tensor strides and memory layouts correctly
- Must use Triton JIT compilation for kernel definition
- Should leverage Triton's autotuning features for optimization
- Kernel must apply GELU activation to the result before returning

Scoring (0-100)
---------------
Performance is measured against baseline implementations:

```
geometric_mean_speedup = geometric_mean(answer_times / baseline_times)
raw_score = min(geometric_mean_speedup, 3.0)  # Cap at 3x speedup
score = (raw_score - 1.0) / 2.0 * 100  # Map 1x-3x to 0-100
```

- 0 points = No speedup (1x baseline performance)
- 50 points = 2x speedup over baseline
- 100 points = 3x+ speedup over baseline

Evaluation Details (near-tile variant)
-------------------------------------
- Shapes clustered around tile boundaries (tile M,N=128, K=64), including +/-1 and +7:
  - M in {127,128,129,135, 255, 385, 633}
  - N in {127,128,129,135, 257, 383, 643}
  - K in {63,64,65,71, 129, 191, 325}
- Only positive dimensions up to 8192 are included; Cartesian product filtered to limits
- Correctness verified with tolerance: rtol=1e-2, atol=5e-3
- Performance measured using median execution time
- Requires CUDA backend and GPU support

Implementation Notes for Solution Authors
----------------------------------------
- Triton `tl.arange(0, BLOCK_*)` requires the range to be a power of two. Choose `BLOCK_M`, `BLOCK_N`, and especially `BLOCK_K` from powers of two (e.g., 32/64/128/256) to avoid compilation errors.
- Return tensor dtype must match input dtype (fp16/bf16/fp32). Accumulate in fp32 inside the kernel, but allocate the output with `dtype=a.dtype` to pass correctness checks.
- Provide a `Solution.solve()` that returns a static code string via `{ "code": python_source }`. Avoid reflection-based approaches (e.g., `inspect.getsource`) as modules are imported under different names during evaluation.
- Respect arbitrary input strides; compute element-wise strides and use masked loads/stores for tail tiles.
- Autotuning: include strides in the autotune key (e.g., `a_stride_am`, `a_stride_ak`, `b_stride_bk`, `b_stride_bn`) to ensure correct kernel specialization across layouts.
- Recommended tile sets to cover near-tile cases:
  - `BLOCK_M/N`: {64, 128, 256}
  - `BLOCK_K`: {32, 64, 128} (avoid non-powers like 80)