[Feature Request] [Pass] DMA-aware MultiVersionBuffer

[firefrogliu666]

Required prerequisites

• I have searched the Issue Tracker that this hasn't already been reported. (comment there if it has.)

Motivation

The SUNMMIO ZPU has an explicit DMA engine and a single control stream (No thread/warp schedualing). It depends on the compiler-generated pipeline to overlap DMA operations and compute rather than implicit warp scheduling.

As for the MultiVersionBuffer pass, we need to remove Thread/Warp Logic from MultiVersionBuffer. The MultiVersionBuffer pass itself has minimal thread-specific code, but we need to ensure:

• The buffer versioning logic in MultiVersionBufferRewriter works without thread assumptions
• The producer/consumer role detection is still valid for our NPU's DMA operations

Solution

No response

Alternatives

No response

Additional context

No response

[yangsmallxiao]

1. The core logic in the original code is as follows:
   1. Identify the buffers to be analyzed: scoped_buffer.
   2. Determine the role of each statement (stmt).
   3. Based on these roles, inspect the scoped_buffer to identify buffers that require multiple versions (versioned_buffer), specifically when:
      • A buffer is used by both producer and consumer statements, or
      • The first write to the buffer occurs before the last read.
   4. Rewrite buffer allocations according to the identified versioned_buffer.
   5. Rewrite buffer accesses according to the identified versioned_buffer.
2. Our specific scenario differs primarily in two aspects:
   • How DMA operations are handled, which deviates from the original approach.
   • How shared_buffer is identified, which uses different criteria.
3. Corresponding modifications to the original steps are as follows:
   1. Revise the identification logic for scoped_buffer, primarily by updating the judgment based on buffer.scope.
   2. Modify the role inference procedure:
      • Replace the detection of tmd_load with recognition of dma_copy().
      • Manually infer from function arguments whether the DMA transfer is from global memory to shared memory or involves other memory spaces.
   3. Refine the logic for identifying statements with the "both" (producer and consumer) role, mainly to maintain compatibility with the updated buffer.scope semantics.
   4. No changes required.
   5. No changes required.