[Bug] [RISC-V RVV] floor operator performance regression

[yanyanyanggg]

Issue: [RISC-V RVV] floor operator performance regression

Description

The floor operator shows performance degradation with the RISC‑V Vector (RVV) extension, achieving only 0.521× the performance of the scalar implementation. This suggests inefficient vectorization for the floor operation.

Steps to Reproduce

1. Generate the floor 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_floor(params, set_dir=None, platform="rv"):
    data = relay.var("data",
                     shape=(params["batch"], params["channels"],
                            params["input_height"], params["input_width"]),
                     dtype=params["dtype"])
    floor_op = relay.floor(data)
    export_op(floor_op, params["op_name"], [data], params, set_dir=set_dir)

Performance Data

• RV execution time: 8.891440 ms
• RVV execution time: 17.061200 ms
• Acceleration ratio (RV/RVV): 0.521 (RVV is ~1.9× 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 floor on ~1.7M elements

Expected Behavior

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

Additional Context

• The floor operation is applied elementwise to a tensor of ~1.7M elements.
• While the regression is less severe than for other operators (sum, log, etc.), it still represents a significant performance degradation for a simple arithmetic operation.
• This suggests that the current RVV implementation of floor may be using suboptimal vector instructions or inefficient vector length management.