Skill: high-performance-gpu-operator-development

[Case 2]

• Initial Environment: A cloud-based GPU development container with the ROCm or CUDA toolchain and a reference implementation of a fused Conv2d-ReLU-MaxPool layer in PyTorch. The environment includes a suite of unit tests with randomized seeds to verify numerical stability.
• Real Question: Implement a fused native C++/CUDA kernel for the forward pass of a sequential module (Convolution -> ReLU -> MaxPool) that outperforms the eager-mode baseline.
• Real Trajectory: The agent analyzes the reference to identify spatial dependencies and constants. It implements a specialized kernel using shared memory to cache input tiles, minimizing the need to re-read data from high-latency global memory for the convolution. It unrolls the inner loops and fuses the activation and pooling logic directly into the tile-writing stage, preventing the creation of intermediate feature map tensors. The agent then runs the profiler to confirm that DRAM throughput is maximized and register spilling is minimized.
• Real Answer: A fused native kernel where convolution, ReLU, and Max-Pooling are performed in one pass, significantly reducing the memory overhead of the neural network layer.
• Why this demonstrates the capability: This illustrates operator fusion, a key high-performance skill. By managing shared memory and unrolling loops, the agent reduces the global memory round-trips that typically slow down sequential AI operators.

Pipeline Execution Instructions

To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:

1. Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds: references/EXPLORATION.md

2. Phase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: references/SELECTION.md

3. Phase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: references/SYNTHESIS.md