Transformer Models Analysis

	A	B	C	D	E	F	G	H	I	J	K	L	M	N	O	P	Q	R	S	T	U	V	W	X	Y	Z	AA	AB	AC	AD	AE	AF	AG	AH	AI
1		V2 ops in red			Optimized: model optimized by ONNX Runtime before performing the inference (before webnn graph building as well)
2	Model Operations (ONNX)	W3C Spec	TOSA	Stable HLO	Op Usage Count of Optimized models	stable-diffusion-v1-5								segment anything		whisper-tiny.en								t5-small				llama-2-7b		gpt2		m2m100_418M
3						unet (fp32)		text_encoder (fp32)		vae_encoder (fp32)		vae_decoder (fp32)		decoder (fp32)		encoder (fp32)		encoder (int8)		decoder (fp32)		decoder (int8)		encoder (fp32)		decoder (fp32)		decoder (fp16)		decoder (fp32)		encoder (fp32)		decoder (fp32)
4						Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized	Original	Optimized
5	Abs	abs			0																			1	0
6	Add	add			15	361	345	160	160	39	38	49	49	68	66	53	53	55	55	80	79	80	79	35	32	52	49	737	226	123	111	160	160	257	256
7	ArgMax	argMax	tosa.argmax		2			1	1					1	1
8	Cast	cast	tosa.cast	stablehlo.convert	9	514	1	3	0	6	0	2	0	54	7			24	24	4	0	41	41	24	1	38	0	706	809	38	1	7	4	10	4
9	Clip	clamp			1					1	1
10	Concat	concat			3	303	14	124	0	10	0	6	0	48	6	32	0	20	0	76	0	50	0	25	0	54	0	642	130	148	0	123	0	246	0
11	Constant	constant			0	2247	0	379	0	142	0	134	0	360	0	120	0			235	0			156	0	295	0	7526	0	1215	0	344	0	659	0
12	ConstantOfShape				0	1	0	2	0	4	0			5	0					2	0	1	0	1	0	6	0	512	0	12	0	1	0	3	0
13	Conv	conv2d			5	98	98			28	28	36	36	3	3	2	2
14	ConvInteger	convInteger			1													2	2
15	ConvTranspose	convTranspose2d			1									2	2
16	Cos	cos			2	1	1							2	1
17	CumSum				2																											1	1	1	1
18	DequantizeLinear	dequantizeLinear			1																	1	1
19	Div	div			13	358	64	25	25	12	0	15	0	36	15	15	15	17	17	17	17	17	17	15	13	21	19	97	97	37	37	25	25	37	37
20	DynamicQuantizeLinear	dynamicQuantizeLinear			2													14	14			26	26
21	Equal	equal	tosa.equal	stablehlo.compare EQ	3	1	0							11	5					1	0	1	0					512	0			2	1	3	1
22	Erf	erf	tosa.erf		6	16	16							4	4	6	6	6	6	4	4	4	4
23	Exp	exp			1					1	1
24	Expand	expand	tosa.tile(a, tosa.div(b.shape, a.shape))	stablehlo.broadcast_in_dim?	5	1	1							6	1					1	0	1	0			1	0	640	128			1	1	2	1
25	Flatten	flattenTo2d			2			1	1					2	2
26	Floor	floor			1													2	2
27	Gather	gather	tosa.gather	stablehlo.gather? (much more complicated)	10	465	0	57	2	17	0	12	0	117	9	8	0	8	0	24	1	24	1	10	1	19	1	705	65	197	1	43	2	80	2
28	Gemm	gemm			3	24	24							15	15															48	48
29	Greater	greater	tosa.greater	stablehlo.compare GT	0																			1	0
30	Identity	identity	tosa.identity		1	129	0	202	0	8	0	6	0							1	1											24	0	36	0
31	InstanceNormalization	meanVarianceNormalization		--- (nearest is stablehlo.batch_norm_inference)	3	61	61			22	22	30	30
32	LayerNormalization	meanVarianceNormalization		--- (nearest is stablehlo.batch_norm_inference)	1									9	9
33	Less	lesser		stablehlo.compare LT	0															1	0	1	0	1	0	1	0							1	0
34	LessOrEqual	lesserOrEqual		stablehlo.compare LE	0																					1	0
35	Log	log			0																			1	0	1	0
36	MatMul	matmul			15	224	224	96	96	6	6	6	6	49	48	32	32	12	12	57	57	16	16	48	48	85	85	289	289	25	25	96	96	169	169
37	MatMulInteger	matmulInteger			2													20	20			41	41
38	Min	min			0																			1	0	2	0
39	Mul	mul			15	435	237	86	61	50	47	62	61	38	21	29	25	75	69	38	29	119	111	29	14	43	20	994	385	74	74	52	39	89	63
40	Neg	neg			0																					1	0
41	Not	logicalNot	tosa.logical_not	stablehlo.not	3									1	1																	1	1	1	1
42	Pad	pad			1					3	3
43	Pow	pow			14	48	48	25	25					3	3	9	9	9	9	13	13	13	13	13	13	19	19	65	65	49	37	25	25	37	37
44	RandomNormalLike				1					1	1
45	Range	fillSequence		stablehlo.iota (lacks start and step)	0			1	0					2	0					1	0	1	0	1	0	2	0	512	0	1	0			1	0
46	ReduceMean	reduceMean			13	96	96	50	50					6	6	18	18	18	18	26	26	26	26	13	13	19	19	65	65	50	50	50	50	74	74
47	Relu	relu			5									12	12									6	6	6	6					12	12	12	12
48	Reshape	reshape	tosa.reshape	stablehlo.reshape	15	411	410	123	123	56	50	66	66	40	37	32	32	34	34	75	73	75	73	25	25	49	49	768	384	148	147	124	123	246	243
49	Resize	resample2d			3	3	3					3	3	2	2
50	ScatterND				1																							128	128
51	Shape				0	527	0	55	0	39	0	42	0	116	0	8	0	4	0	26	0	15	0	9	0	20	0	1536	0	280	0	42	0	82	0
52	Sigmoid	sigmoid			5	68	47	12	12	21	21	29	29															32	32
53	Sin	sin			2	1	1							2	1
54	Slice	slice			4	34	34	1	0	5	2			7	3					3	0	2	0					832	260	109	0			2	0
55	Softmax	softmax			15	32	32	12	12	1	1	1	1	7	7	4	4	4	4	8	8	8	8	6	6	12	12	32	32	12	12	12	12	24	24
56	Split	split			2									4	1															12	12
57	Sqrt	sqrt	tosa.reciprocal(tosa.rsqrt) or tosa.pow	stablehlo.sqrt	13	48	48	25	25					3	3	9	9	9	9	13	13	13	13	13	13	19	19	65	65	25	25	25	25	37	37
58	Squeeze	squeeze			0															2	0	1	0							85	0			2	0
59	Sub	sub			12	48	48	25	25					6	5	9	9	9	9	13	13	13	13	2	1	3	0	64	64	38	26	26	26	38	38
60	Tanh	tanh			1																									12	12
61	Tile				0																					1	0
62	Transpose	transpose			15	192	192	60	60	9	6	6	6	32	30	21	21	21	21	40	40	40	40	25	24	61	60	128	128	60	60	60	60	121	120
63	Trilu	triangularMatrix			0			1	0
64	Unsqueeze				7	926	45	222	0	19	0	15	0	107	10	44	0	12	0	122	0	33	0	30	2	70	0	1440	66	285	2	212	2	424	2
65	Where	elementwiseIf	tosa.select	stablehlo.select	3	1	0							5	0					2	0	2	0	1	0	1	0	512	0	12	12	2	1	4	1
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100