cpu: rv64: add rvv inner product implementation using rvv intrinsics

[zhangjian29]

Description

This PR introduces optimized inner product (ip) primitive for RV64 architectures using RVV (RISC-V Vector) intrinsics. The current rvv_inner_product implementation focus on 2-D plain inputs layouts.

This initial version provides:

1. Supported memory layouts: plain, dense, no inner blocks
2. Supported dst data types: f32, s32, s8, u8
3. Support only 2D source/weights with data type combinations of f32:f32, s8:s8 and u8:s8
4. Support bias computation

Implementation Details

• Memory Reorder: Reorder src and weight if they are non-continuous​ in memory using scratchpad method
• Seperate Computations: Seperate inner product computations into compute_ip_rvv and finalize_ip_acc:
  • compute_ip_rvv: Do the vector compuation using different RVV intrinsics for various src/weight data type combinations
  • finalize_ip_acc: After vector compuation, get the scalar result and add bias (will do post-ops here in the future), then clamp dst and convert to the required data type

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?

Performance improvements

All experiments are performed on a SG2044 platform including:

• Test Cases: benchdnn inputs of alexnet / resnet_50 / rnn_t / transformer_lt
• Test Dtypes: f32:f32:f32 / s8:s8:f32 / u8:s8:s32
• Test Args: --mode=p --dir=FWD_B

We draw comparisons among 1st baseline method of scalar implementation, 2nd baseline method of auto vectoration by compiler,and our method of RVV intrinsic implementation.

1. Scalar: ref_gemm_inner_product implementation compiled by gcc 14.2 with -march=rv64gc -O3
2. Auto Vectorization: ref_gemm_inner_product implementation compiled by gcc 14.2 with -march=rv64gcv -O3 -ftree-vectorize
3. RVV Intrinsic: Our rvv_inner_product implementation compiled by gcc 14.2 with -march=rv64gc -O3

Results

Averagely, runtimes using rvv instrinsics have obtained a 1.49x speep up compared to those with scalar method, and a 1.28x speep up compared to those with compiler auto vectorization.

In the best test case of --dtu8:s8:s32 mb4ic1024oc4096n"Transformer_lt:Decoder_MM_7*240", rvv intrinsics have a 9.55x speep up compared to scalar method, and a 17.01x speep up compared to compiler auto vectorization.

Detailed results are as follows:

Methods	Data types	Cases	Runtime (ms)	Speedups
scalar	f32:f32:f32	alexnet	1785.14	1.00
scalar	f32:f32:f32	resnet_50	26.2	0.70
scalar	f32:f32:f32	rnn_t	314.22	1.54
scalar	f32:f32:f32	transformer_lt	159.75	0.51
scalar	s8:s8:f32	alexnet	1860.67	1.32
scalar	s8:s8:f32	resnet_50	46.22	3.63
scalar	s8:s8:f32	rnn_t	493.47	2.43
scalar	s8:s8:f32	transformer_lt	746.73	4.28
scalar	u8:s8:s32	alexnet	2329.33	1.46
scalar	u8:s8:s32	resnet_50	68.54	4.80
scalar	u8:s8:s32	rnn_t	439.22	1.91
scalar	u8:s8:s32	transformer_lt	900.87	5.08
auto vectorization	f32:f32:f32	alexnet	476.65	0.27
auto vectorization	f32:f32:f32	resnet_50	33.74	0.91
auto vectorization	f32:f32:f32	rnn_t	246.08	1.21
auto vectorization	f32:f32:f32	transformer_lt	178.69	0.57
auto vectorization	s8:s8:f32	alexnet	2551.34	1.81
auto vectorization	s8:s8:f32	resnet_50	50.26	3.95
auto vectorization	s8:s8:f32	rnn_t	812.71	4.01
auto vectorization	s8:s8:f32	transformer_lt	783.0	4.48
auto vectorization	u8:s8:s32	alexnet	1412.33	0.88
auto vectorization	u8:s8:s32	resnet_50	46.65	3.27
auto vectorization	u8:s8:s32	rnn_t	488.94	2.12
auto vectorization	u8:s8:s32	transformer_lt	799.07	4.51
rvv intrinsic	f32:f32:f32	alexnet	1794.01	-
rvv intrinsic	f32:f32:f32	resnet_50	37.19	-
rvv intrinsic	f32:f32:f32	rnn_t	203.82	-
rvv intrinsic	f32:f32:f32	transformer_lt	314.39	-
rvv intrinsic	s8:s8:f32	alexnet	1409.76	-
rvv intrinsic	s8:s8:f32	resnet_50	12.72	-
rvv intrinsic	s8:s8:f32	rnn_t	202.84	-
rvv intrinsic	s8:s8:f32	transformer_lt	174.67	-
rvv intrinsic	u8:s8:s32	alexnet	1600.82	-
rvv intrinsic	u8:s8:s32	resnet_50	14.28	-
rvv intrinsic	u8:s8:s32	rnn_t	230.54	-
rvv intrinsic	u8:s8:s32	transformer_lt	177.26	-