[Bug] [RISC-V RVV] Performance Degradation: ReLU activation slower with vector extension

[yanyanyanggg]

Issue: [RISC-V RVV] Performance Degradation: ReLU activation slower with vector extension

Description

The ReLU (rectified linear unit) operator shows significant performance degradation with the RISC‑V Vector (RVV) extension. The acceleration ratio is 0.337, meaning the RVV version is about 3× slower than the scalar implementation. This is unexpected for a simple elementwise operation that should benefit greatly from vectorization.

Steps to Reproduce

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

Performance Data

• RV execution time: 7.945310 ms
• RVV execution time: 23.579300 ms
• Acceleration ratio (RV/RVV): 0.337 (RVV is ~3× 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 ReLU on ~1.7M elements

Expected Behavior

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

Additional Context

• The ReLU operation is applied elementwise to a tensor of ~1.7M elements.
• The severe performance regression (3× slower) is particularly surprising for such a simple operation that should be a perfect candidate for vectorization.
• This issue is part of a broader pattern where multiple operators (sum, log, relu, bias_add, sqrt, etc.) show significant performance degradation with RVV, suggesting a potential systemic issue in TVM's RVV code generation or optimization.