cpu: rv64: add support for f32 layer normalization using RVV

[krishnasai-mcw]

Description

This PR introduces optimized layer normalization primitive for RV64 architectures using RVV (RISC-V Vector) intrinsics.

Features and Limitations

Supported:

• Data type: f32
• Input: plain and dense layouts, with the last dimension contiguous
• Stats modes:
  • Global stats (mean and variance provided externally)
  • Calculate stats (kernel computes mean and variance internally)
• Scale & shift
• Forward propagation - FWD_I & FWD_D

Not supported:

• Post-ops
• RMS layer normalization
• Backward propagation

Implementation Details

The implementation uses an LMUL=1 configuration with 4-way loop unrolling to maximize throughput while maintaining register pressure. For statistics calculation, the kernel uses double-precision (f64) accumulators via widening operations to maintain numerical accuracy across large reduction operations. The normalization pass applies the standard formula ((x - mean) / std) × scale + shift using fused multiply-accumulate instructions to efficiently compute the final output.

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

BenchDNN test log: test_lnorm_all.log

Performance improvements

• Have you submitted performance data that demonstrates performance improvements?

Performance Results

Test Platform: Banana Pi F3

Test Case	Vector	Scalar	Speedup
option_set_all	86.46	1055.51	12.21
option_set_fwks_ext_gpu	62.36	765.01	12.27
shapes_ci	2.67	29.36	11.00

[zhangjian29]

Hi @krishnasai-mcw ,

I noticed that you're using f64 data type. Do you think it is necessary to add a check to ensure the platform supports f64?

[krishnasai-mcw]

Hi @krishnasai-mcw ,

I noticed that you're using f64 data type. Do you think it is necessary to add a check to ensure the platform supports f64?

Hi @zhangjian29 ,
No extra check for f64 is needed. The RISC-V G extension itself includes the D extension, which provides full double-precision (f64) support. In our CMake setup, we already target G by default, ensuring f64 is supported by default.

Thanks