yanyanyanggg opened a new issue, #18561:
URL: https://github.com/apache/tvm/issues/18561
### Issue: [RISC-V RVV] Performance Issue: log operator slower on RVV
#### Description
The log operator exhibits performance regression with the RISC‑V Vector
(RVV) extension enabled. The acceleration ratio is 0.328, indicating the RVV
implementation is approximately 3× slower than the scalar RV version. This is
unexpected for an elementwise mathematical operation that should benefit from
vectorization.
#### Steps to Reproduce
1. Generate the log operator with the following configuration:
```python
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:
```python
def export_log(params, set_dir=None, platform="rv"):
data = relay.var("data",
shape=(params["batch"], params["channels"],
params["input_height"], params["input_width"]),
dtype=params["dtype"])
log_op = relay.log(data)
export_op(log_op, params["op_name"], [data], params, set_dir=set_dir)
```
#### Performance Data
- **RV execution time**: 13.393300 ms
- **RVV execution time**: 40.848000 ms
- **Acceleration ratio (RV/RVV)**: 0.328 (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
- **Input shape**: (14, 23, 67, 99) ≈ 1.7M elements
#### Expected Behavior
RVV vectorization should provide a performance improvement over the scalar
RV baseline for elementwise mathematical operations like log.
#### Additional Context
- The log operation is applied elementwise to a tensor of ~1.7M elements.
- This severe performance regression suggests inefficient vector code
generation or suboptimal use of RVV instructions for mathematical functions.
- Similar regressions are observed across multiple operators (sum, relu,
bias_add, sqrt, etc.), indicating a potential systemic issue with RVV
vectorization in TVM.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
