[Bug] [RISC-V RVV] round operator shows suboptimal vectorization

[yanyanyanggg]

Issue: [RISC-V RVV] round operator shows suboptimal vectorization

Description

The round operator performs worse with the RISC‑V Vector (RVV) extension, achieving only 0.547× the performance of the scalar implementation. This indicates inefficient vectorization for rounding operations.

Steps to Reproduce

1. Generate the round operator with the following configuration:

params = {
    "dtype": "float32",
    "batch": 14,
    "channels": 23,
    "input_height": 67,
    "input_width": 99
}

2. Export the operator to two targets:

   • RV target (scalar, without vector extension):

     llvm -mtriple=riscv64-linux-gnu -mcpu=generic-rv64 -mabi=lp64d -mattr=+64bit,+m,+a,+f,+d,+c
     

   • RVV target (with vector extension):

     llvm -mtriple=riscv64-linux-gnu -mcpu=generic-rv64 -mabi=lp64d -mattr=+64bit,+m,+a,+f,+d,+c,+v
     

3. Run performance measurement on both targets.

Operator definition code:

def export_round(params, set_dir=None, platform="rv"):
    data = relay.var("data",
                     shape=(params["batch"], params["channels"],
                            params["input_height"], params["input_width"]),
                     dtype=params["dtype"])
    round_op = relay.round(data)
    export_op(round_op, params["op_name"], [data], params, set_dir=set_dir)

Performance Data

• RV execution time: 7.314920 ms
• RVV execution time: 13.376600 ms
• Acceleration ratio (RV/RVV): 0.547 (RVV is ~1.8× slower)

Environment Information

• TVM version: 0.19.0
• LLVM version: [Please provide: llvm-config --version]
• Hardware: Spacemit K1‑X bit‑brick board
• CPU: Spacemit X60 (8 cores, 1.6 GHz)
• ISA: rv64imafdcv (with vector extensions)
• Memory: 7.6 GB
• OS: Bianbu 2.2, Linux kernel 6.6.63
• Operation: Elementwise rounding on ~1.7M elements

Expected Behavior

RVV vectorization should provide a performance improvement over the scalar RV baseline for elementwise operations like round.

Additional Context

• The round operation is applied elementwise to a tensor of ~1.7M elements.
• The performance regression is significant and suggests that the vectorized implementation of round may be using inefficient instructions or suboptimal vector length.
• This is part of a pattern where multiple elementwise operations (including floor, round, etc.) show performance degradation with RVV, indicating a potential systemic issue in the vectorization of these operations.