openai/openai-gemm

Open single and half precision gemm implementations

397

14 commits

updated Apr 2, 2023

See the code

README

Status: Archive (code is provided as-is, no updates expected)

openai-gemm

Open single and half precision gemm implementations. The main speedups over cublas are with small minibatch and in fp16 data formats.

Quick Install

The demonstration code currently depends on Nervana neon:

git clone [email protected]:NervanaSystems/neon.git
cd neon
make
. .venv/bin/activate

Clone and run this repo:

git clone [email protected]:openai/openai-gemm.git

Run the benchmark:
./benchmark.py

Run the unit test:
./test.py

DeepBench on Pascal TITAN X

( https://github.com/baidu-research/DeepBench )

M	N	K	Op	OpenAI_32	cuBLAS_32	ratio_32	OpenAI_16	cuBLAS_16	ratio_16
16	1760	1760	NN	2557	2195	1.2	3507	346	10.1
32	1760	1760	NN	5010	1128	4.4	6814	526	13.0
64	1760	1760	NN	6486	4112	1.6	8235	2801	2.9
128	1760	1760	NN	7068	6931	1.0	9400	5307	1.8
7000	1760	1760	NN	9968	9584	1.0	10515	9807	1.1
16	2048	2048	NN	2569	1516	1.7	3619	242	15.0
32	2048	2048	NN	5034	1356	3.7	6576	606	10.8
64	2048	2048	NN	6636	2815	2.4	8285	3241	2.6
128	2048	2048	NN	7316	6373	1.1	9066	5334	1.7
7000	2048	2048	NN	10081	9900	1.0	11275	9948	1.1
16	2560	2560	NN	2718	1312	2.1	4312	251	17.2
32	2560	2560	NN	5370	1660	3.2	7525	749	10.0
64	2560	2560	NN	7331	2687	2.7	8436	951	8.9
128	2560	2560	NN	8007	5238	1.5	9277	6123	1.5
7000	2560	2560	NN	10282	10131	1.0	11027	9974	1.1
16	4096	4096	NN	2695	1110	2.4	4442	266	16.7
32	4096	4096	NN	5266	2264	2.3	7723	758	10.2
64	4096	4096	NN	6942	3922	1.8	8904	1055	8.4
128	4096	4096	NN	8127	5686	1.4	9711	5681	1.7
7000	4096	4096	NN	10462	10082	1.0	11152	9991	1.1
16	1760	1760	NT	1719	1095	1.6	2692	290	9.3
32	1760	1760	NT	3316	1312	2.5	5068	447	11.3
64	1760	1760	NT	5247	1955	2.7	7621	1797	4.2
128	1760	1760	NT	6720	3393	2.0	8886	3342	2.7
7000	1760	1760	NT	9341	8513	1.1	10085	9635	1.0
16	2048	2048	NT	2442	1231	2.0	3641	299	12.2
32	2048	2048	NT	4801	1251	3.8	5849	468	12.5
64	2048	2048	NT	6317	1967	3.2	7825	3128	2.5
128	2048	2048	NT	7176	5041	1.4	8616	4843	1.8
7000	2048	2048	NT	9975	9173	1.1	10741	9560	1.1
16	2560	2560	NT	1834	1208	1.5	3154	297	10.6
32	2560	2560	NT	3610	1436	2.5	5418	584	9.3
64	2560	2560	NT	6083	2815	2.2	8331	1042	8.0
128	2560	2560	NT	7702	3246	2.4	8857	5259	1.7
7000	2560	2560	NT	9257	7829	1.2	10659	9548	1.1
16	4096	4096	NT	2546	1297	2.0	4164	309	13.5
32	4096	4096	NT	4992	2290	2.2	8156	775	10.5
64	4096	4096	NT	6746	4157	1.6	8429	1381	6.1
128	4096	4096	NT	7843	5425	1.4	9298	5527	1.7
7000	4096	4096	NT	9925	6879	1.4	10630	9784	1.1
7133	1760	1760	TN	9752	10186	1.0	10517	8912	1.2
7133	2048	2048	TN	10485	10319	1.0	10674	9608	1.1
7133	2560	2560	TN	10743	11057	1.0	11195	10059	1.1
7133	4096	4096	TN	10384	10290	1.0	10980	10558	1.0
9124	5124	1760	NN	9920	9480	1.0	10580	9743	1.1
9124	5124	2048	NN	10008	9415	1.1	10602	9796	1.1
9124	5124	2560	NN	9925	9426	1.1	10586	9850	1.1
9124	5124	4096	NN	9982	9489	1.1	10580	9472	1.1
9124	5124	1760	NT	9093	3497	2.6	9302	8692	1.1
9124	5124	2048	NT	9506	6512	1.5	9506	8883	1.1
9124	5124	2560	NT	8704	3364	2.6	9855	7733	1.3
9124	5124	4096	NT	9733	6109	1.6	10278	8760	1.2
8457	35	1760	NN	3343	1020	3.3	3841	736	5.2
8457	35	2048	NN	3419	1996	1.7	4782	803	6.0
8457	35	2560	NN	3415	1072	3.2	3868	789	4.9
8457	35	4096	NN	3743	2009	1.9	4741	804	5.9
8457	35	1760	NT	3574	1970	1.8	4176	1243	3.4
8457	35	2048	NT	4564	3069	1.5	4818	1255	3.8
8457	35	2560	NT	3598	2062	1.7	3597	1135	3.2
8457	35	4096	NT	4311	2990	1.4	4927	1303	3.8
16	7680	2560	NN	2683	718	3.7	4449	289	15.4
32	7680	2560	NN	5304	3660	1.4	7837	979	8.0
64	7680	2560	NN	7311	4979	1.5	9310	1274	7.3
128	7680	2560	NN	7931	6109	1.3	9390	6591	1.4
16	7680	2560	NT	1885	1191	1.6	3401	290	11.7
32	7680	2560	NT	3731	1808	2.1	6373	1004	6.3
64	7680	2560	NT	6274	3509	1.8	8809	1655	5.3
128	7680	2560	NT	7957	2988	2.7	9246	4695	2.0
16	3072	1024	NN	2277	1295	1.8	3373	282	12.0
32	3072	1024	NN	4494	1798	2.5	6011	807	7.4
64	3072	1024	NN	6272	3046	2.1	6790	917	7.4
128	3072	1024	NN	7364	5436	1.4	7768	5749	1.4
16	3072	1024	NT	2285	1077	2.1	3439	244	14.1
32	3072	1024	NT	4597	1540	3.0	5645	677	8.3
64	3072	1024	NT	6392	2969	2.2	7555	1204	6.3
128	3072	1024	NT	7460	5058	1.5	8586	5535	1.6
7435	3072	1024	TN	9829	8804	1.1	10123	9365	1.1
5481	7680	2560	TN	9448	9309	1.0	9466	9394	1.0

DeepBench on DGX1 (P100)

Note that the OpenAI kernels do not yet implement fp16x2 instructions. Even still it seems the current cublas hgemm implentation is only good for large dimensions. There are also accuracy considerations when accumulating large reductions in fp16.

M	N	K	Op	OpenAI_32	cuBLAS_32	ratio_32	OpenAI_16	cuBLAS_16	ratio_16
16	1760	1760	NN	2595	2048	1.3	2935	463	6.3
32	1760	1760	NN	4963	864	5.7	5766	895	6.4
64	1760	1760	NN	7565	3909	1.9	7760	1711	4.5
128	1760	1760	NN	8140	6053	1.3	8422	4089	2.1
7000	1760	1760	NN	9653	8722	1.1	9617	16143	0.6
16	2048	2048	NN	2255	1746	1.3	3211	546	5.9
32	2048	2048	NN	4467	1012	4.4	4533	1019	4.4
64	2048	2048	NN	6618	4198	1.6	6591	2018	3.3
128	2048	2048	NN	8059	5921	1.4	7936	4667	1.7
7000	2048	2048	NN	9761	9346	1.0	9910	18715	0.5
16	2560	2560	NN	2883	2108	1.4	4210	685	6.1
32	2560	2560	NN	5701	1279	4.5	5820	1297	4.5
64	2560	2560	NN	8100	6054	1.3	8099	2558	3.2
128	2560	2560	NN	8308	6799	1.2	8790	5901	1.5
7000	2560	2560	NN	9740	9538	1.0	9845	18499	0.5
16	4096	4096	NN	3449	1342	2.6	4299	1069	4.0
32	4096	4096	NN	6863	2045	3.4	6907	2103	3.3
64	4096	4096	NN	8404	4059	2.1	8248	4183	2.0
128	4096	4096	NN	8224	8039	1.0	8853	8669	1.0
7000	4096	4096	NN	9818	9519	1.0	10011	18588	0.5
16	1760	1760	NT	2579	1324	1.9	2763	428	6.4
32	1760	1760	NT	5089	878	5.8	5382	857	6.3
64	1760	1760	NT	7501	3017	2.5	7695	1695	4.5
128	1760	1760	NT	8043	5494	1.5	8192	3426	2.4
7000	1760	1760	NT	9477	7571	1.3	9355	16113	0.6
16	2048	2048	NT	2267	1276	1.8	3171	504	6.3
32	2048	2048	NT	4484	1026	4.4	4489	1009	4.4
64	2048	2048	NT	6567	3986	1.6	6551	2018	3.2
128	2048	2048	NT	8019	5825	1.4	7968	4496	1.8
7000	2048	2048	NT	9625	9373	1.0	9713	17878	0.5
16	2560	2560	NT	2870	1460	2.0	4256	638	6.7
32	2560	2560	NT	5614	1299	4.3	5705	1271	4.5
64	2560	2560	NT	8014	4402	1.8	8085	2521	3.2
128	2560	2560	NT	8219	5640	1.5	8240	5137	1.6
7000	2560	2560	NT	9534	9091	1.0	9735	18025	0.5
16	4096	4096	NT	3366	1547	2.2	4354	1047	4.2
32	4096	4096	NT	6714	2055	3.3	6859	2093	3.3
64	4096	4096	NT	8297	3445	2.4	8289	4178	2.0
128	4096	4096	NT	8335	7450	1.1	7911	7973	1.0
7000	4096	4096	NT	9578	9214	1.0	9877	18073	0.5
7133	1760	1760	TN	9704	9267	1.0	9506	15605	0.6
7133	2048	2048	TN	9747	9836	1.0	10012	19110	0.5
7133	2560	2560	TN	9742	9748	1.0	9805	19107	0.5
7133	4096	4096	TN	9807	9733	1.0	10122	19559	0.5
9124	5124	1760	NN	9326	9076	1.0	9631	17496	0.6
9124	5124	2048	NN	9414	9054	1.0	9602	17523	0.5
9124	5124	2560	NN	9353	9041	1.0	9698	17380	0.6
9124	5124	4096	NN	9370	9051	1.0	9689	17617	0.5
9124	5124	1760	NT	9124	8746	1.0	9524	16777	0.6
9124	5124	2048	NT	9294	8817	1.1	9641	16935	0.6
9124	5124	2560	NT	9221	8499	1.1	9637	16820	0.6
9124	5124	4096	NT	9270	8961	1.0	9568	17080	0.6
8457	35	1760	NN	3301	2233	1.5	4505	3154	1.4
8457	35	2048	NN	3265	3066	1.1	4501	3335	1.3
8457	35	2560	NN	3127	2300	1.4	4516	3135	1.4
8457	35	4096	NN	3257	3272	1.0	4729	3485	1.4
8457	35	1760	NT	4563	3142	1.5	4612	2998	1.5
8457	35	2048	NT	4554	3202	1.4	4601	3109	1.5
8457	35	2560	NT	4567	3144	1.5	4654	3039	1.5
8457	35	4096	NT	4353	3415	1.3	4457	3257	1.4
16	7680	2560	NN	3668	1200	3.1	5020	1236	4.1
32	7680	2560	NN	7245	3385	2.1	7519	2465	3.1
64	7680	2560	NN	8440	5210	1.6	8349	4910	1.7
128	7680	2560	NN	8765	4872	1.8	9131	11349	0.8
16	7680	2560	NT	3229	1515	2.1	5032	1157	4.3
32	7680	2560	NT	6640	2721	2.4	6810	2307	3.0
64	7680	2560	NT	8282	5113	1.6	8362	4494	1.9
128	7680	2560	NT	8763	4646	1.9	8617	9159	0.9
16	3072	1024	NN	2929	1717	1.7	3335	750	4.4
32	3072	1024	NN	5801	1399	4.1	6116	1420	4.3
64	3072	1024	NN	6958	4340	1.6	6923	2814	2.5
128	3072	1024	NN	8047	6492	1.2	7769	6302	1.2
16	3072	1024	NT	2990	1068	2.8	3384	705	4.8
32	3072	1024	NT	5834	1429	4.1	6021	1411	4.3
64	3072	1024	NT	6921	3500	2.0	6893	2819	2.4
128	3072	1024	NT	7918	6034	1.3	7876	5760	1.4
7435	3072	1024	TN	9367	9391	1.0	9559	17234	0.6
5481	7680	2560	TN	9672	9520	1.0	9967	18832	0.5

Significant stargazers

Valentin Churavy

674 followers · starred Apr 2017

Jerry Zhang

295 followers · starred May 2017

Yuxin Wu

4,004 followers · starred Nov 2016

Xiaoyu Zhang

2,858 followers · starred Jun 2021

Similar repos

wangsiping97/FastGEMV

High-speed GEMV kernels, at most 2.7x speedup compared to pytorch baseline.

130

ornith-ai/CUDA-L2

CUDA-L2: Surpassing cuBLAS Performance for Matrix Multiplication through Reinforcement Learning

484

aramesh10/OpenGEMM

GEMM kernels for B200s

baidu-research/DeepBench

Benchmarking Deep Learning operations on different hardware

1,109

ooples/AiDotNet.Tensors

Pure managed C# tensor library for .NET: hand-written AVX2/AVX-512 SIMD kernels, autodiff and…

deepseek-ai/DeepGEMM-Ascend

DeepGEMM-Ascend: clean and efficient matrix multiplication kernel library for Huawei Ascend NPUs

588

xlite-dev/LeetCUDA

An Open-source book with Modern CUDA learning Notes for Beginners - BF16/FP8/FP4, HGEMM, CuTe,…

12,096

tile-ai/BitBLAS

ReDesign of https://github.com/microsoft/BitBLAS