【KernelGen】Add rsub operator

[Schopenhauer-loves-Hegel]

PR Category

Operator

Type of Change

New Feature

Description

Add rsub operator implementation with Triton kernel.

• Implementation mode: pointwise_dynamic
• Accuracy test: 360/360 passed

Issue

N/A

Progress

• Change is properly reviewed (1 reviewer required, 2 recommended).
• Change is responded to an issue.
• Change is fully covered by a UT.

Performance

torch.bfloat16

Shape	Torch Latency (ms)	Gems Latency (ms)	Speedup
[1073741824]	4.7474	4.6703	1.017
[64, 64]	0.0068	0.0074	0.926
[4096, 4096]	0.0857	0.0857	1.001
[64, 512, 512]	0.0856	0.0857	0.999
[1024, 1024, 1024]	4.7481	4.6625	1.018
[1024, 1]	0.0071	0.0072	0.987
[1024, 16]	0.0074	0.0071	1.050
[1024, 256]	0.0086	0.0086	1.000
[1024, 4096]	0.0286	0.0288	0.992
[1024, 65536]	0.3061	0.3040	1.007
[64, 64, 1]	0.0078	0.0074	1.061
[64, 64, 16]	0.0073	0.0073	1.000
[64, 64, 256]	0.0134	0.0131	1.022
[64, 64, 4096]	0.0856	0.0856	1.000
[64, 64, 65536]	1.1927	1.1854	1.006

torch.float16

Shape	Torch Latency (ms)	Gems Latency (ms)	Speedup
[1073741824]	4.7495	4.6840	1.014
[64, 64]	0.0073	0.0074	0.987
[4096, 4096]	0.0857	0.0856	1.001
[64, 512, 512]	0.0863	0.0855	1.009
[1024, 1024, 1024]	4.7464	4.6970	1.011
[1024, 1]	0.0068	0.0072	0.951
[1024, 16]	0.0074	0.0071	1.050
[1024, 256]	0.0086	0.0093	0.925
[1024, 4096]	0.0278	0.0287	0.968
[1024, 65536]	0.3060	0.3032	1.009
[64, 64, 1]	0.0080	0.0076	1.050
[64, 64, 16]	0.0078	0.0077	1.017
[64, 64, 256]	0.0133	0.0133	1.000
[64, 64, 4096]	0.0856	0.0855	1.001
[64, 64, 65536]	1.1934	1.1858	1.006

torch.float32

Shape	Torch Latency (ms)	Gems Latency (ms)	Speedup
[1073741824]	9.4794	9.3395	1.015
[64, 64]	0.0068	0.0071	0.964
[4096, 4096]	0.1579	0.1578	1.001
[64, 512, 512]	0.1576	0.1575	1.001
[1024, 1024, 1024]	9.4795	9.3306	1.016
[1024, 1]	0.0068	0.0072	0.951
[1024, 16]	0.0071	0.0076	0.925
[1024, 256]	0.0101	0.0110	0.922
[1024, 4096]	0.0474	0.0480	0.987
[1024, 65536]	0.6022	0.5972	1.008
[64, 64, 1]	0.0074	0.0081	0.909
[64, 64, 16]	0.0079	0.0076	1.042
[64, 64, 256]	0.0165	0.0169	0.979
[64, 64, 4096]	0.1578	0.1584	0.996
[64, 64, 65536]	2.3780	2.3496	1.012

Overall: median speedup = 1.001x, mean speedup = 0.996x (45 data points)

---

Generated by auto_gen tool with Claude Code

[Schopenhauer-loves-Hegel]

Closing in favor of grouped PRs to reduce review overhead.

[CLAassistant]

All committers have signed the CLA.

[factnn]

Benchmark Results for rsub operator

float16

Torch (ms)	Gems (ms)	Speedup	Size
14.59	16.11	0.906	[1073741824]
0.007	0.008	0.878	[64,64]
0.085	0.089	0.964	[4096,4096]
0.086	0.089	0.965	[64,512,512]
13.32	17.02	0.782	[1024,1024,1024]
0.007	0.008	0.905	[1024,1]
0.008	0.008	0.967	[1024,16]
0.009	0.010	0.877	[1024,256]
0.028	0.030	0.930	[1024,4096]
0.305	0.312	0.979	[1024,65536]
0.007	0.008	0.873	[64,64,1]
0.008	0.007	1.122	[64,64,16]
0.012	0.014	0.879	[64,64,256]
0.085	0.088	0.975	[64,64,4096]
5.14	7.53	0.683	[64,64,65536]

float32

Torch (ms)	Gems (ms)	Speedup	Size
24.30	30.52	0.796	[1073741824]
0.007	0.008	0.923	[64,64]
0.157	0.159	0.987	[4096,4096]
0.158	0.158	0.996	[64,512,512]
40.02	41.85	0.956	[1024,1024,1024]
0.008	0.008	1.000	[1024,1]
0.007	0.007	0.998	[1024,16]
0.010	0.011	0.922	[1024,256]
0.047	0.048	0.987	[1024,4096]
0.602	0.597	1.008	[1024,65536]
0.007	0.008	0.909	[64,64,1]
0.008	0.008	1.042	[64,64,16]
0.017	0.017	0.979	[64,64,256]
0.158	0.158	0.996	[64,64,4096]
2.38	2.35	1.012	[64,64,65536]

bfloat16

Torch (ms)	Gems (ms)	Speedup	Size
13.41	16.42	0.817	[1073741824]
0.007	0.008	0.878	[64,64]
0.086	0.089	0.966	[4096,4096]
0.086	0.089	0.965	[64,512,512]
13.41	16.42	0.817	[1024,1024,1024]
0.008	0.007	1.169	[1024,1]
0.007	0.008	0.858	[1024,16]
0.009	0.010	0.902	[1024,256]
0.028	0.030	0.925	[1024,4096]
0.306	0.312	0.979	[1024,65536]
0.007	0.007	0.989	[64,64,1]
0.008	0.009	0.953	[64,64,16]
0.012	0.014	0.898	[64,64,256]
0.085	0.089	0.958	[64,64,4096]
4.88	5.29	0.923	[64,64,65536]

Overall: median speedup ~0.96x, mean ~0.94x — performance is on par with PyTorch, as expected for a simple pointwise op.

[tengqm]

Overall lgtm.

[factnn]

UT Results

Branch rebased to upstream/master and tests passed:

python -m pytest tests/ -m rsub --ref cpu -x -q
36 passed