rocm-systems/tests/workloads/invdev/mi200/pmc_dispatch_info.csv

1	Index	KernelName	gpu-id
2	0	__amd_rocclr_fillBufferAligned.kd	0
3	1	void benchmark_func<short, 256, 8u, 0u>(short, short*) [clone .kd]	0
4	2	void benchmark_func<float, 256, 8u, 0u>(float, float*) [clone .kd]	0
5	3	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 0u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
6	4	void benchmark_func<double, 256, 8u, 0u>(double, double*) [clone .kd]	0
7	5	void benchmark_func<__half2, 256, 8u, 0u>(__half2, __half2*) [clone .kd]	0
8	6	void benchmark_func<int, 256, 8u, 0u>(int, int*) [clone .kd]	0
9	7	void benchmark_func<float, 256, 8u, 1u>(float, float*) [clone .kd]	0
10	8	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 1u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
11	9	void benchmark_func<double, 256, 8u, 1u>(double, double*) [clone .kd]	0
12	10	void benchmark_func<__half2, 256, 8u, 1u>(__half2, __half2*) [clone .kd]	0
13	11	void benchmark_func<int, 256, 8u, 1u>(int, int*) [clone .kd]	0
14	12	void benchmark_func<float, 256, 8u, 2u>(float, float*) [clone .kd]	0
15	13	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 2u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
16	14	void benchmark_func<double, 256, 8u, 2u>(double, double*) [clone .kd]	0
17	15	void benchmark_func<__half2, 256, 8u, 2u>(__half2, __half2*) [clone .kd]	0
18	16	void benchmark_func<int, 256, 8u, 2u>(int, int*) [clone .kd]	0
19	17	void benchmark_func<float, 256, 8u, 3u>(float, float*) [clone .kd]	0
20	18	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 3u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
21	19	void benchmark_func<double, 256, 8u, 3u>(double, double*) [clone .kd]	0
22	20	void benchmark_func<__half2, 256, 8u, 3u>(__half2, __half2*) [clone .kd]	0
23	21	void benchmark_func<int, 256, 8u, 3u>(int, int*) [clone .kd]	0
24	22	void benchmark_func<float, 256, 8u, 4u>(float, float*) [clone .kd]	0
25	23	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 4u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
26	24	void benchmark_func<double, 256, 8u, 4u>(double, double*) [clone .kd]	0
27	25	void benchmark_func<__half2, 256, 8u, 4u>(__half2, __half2*) [clone .kd]	0
28	26	void benchmark_func<int, 256, 8u, 4u>(int, int*) [clone .kd]	0
29	27	void benchmark_func<float, 256, 8u, 5u>(float, float*) [clone .kd]	0
30	28	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 5u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
31	29	void benchmark_func<double, 256, 8u, 5u>(double, double*) [clone .kd]	0
32	30	void benchmark_func<__half2, 256, 8u, 5u>(__half2, __half2*) [clone .kd]	0
33	31	void benchmark_func<int, 256, 8u, 5u>(int, int*) [clone .kd]	0
34	32	void benchmark_func<float, 256, 8u, 6u>(float, float*) [clone .kd]	0
35	33	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 6u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
36	34	void benchmark_func<double, 256, 8u, 6u>(double, double*) [clone .kd]	0
37	35	void benchmark_func<__half2, 256, 8u, 6u>(__half2, __half2*) [clone .kd]	0
38	36	void benchmark_func<int, 256, 8u, 6u>(int, int*) [clone .kd]	0
39	37	void benchmark_func<float, 256, 8u, 7u>(float, float*) [clone .kd]	0
40	38	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 7u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
41	39	void benchmark_func<double, 256, 8u, 7u>(double, double*) [clone .kd]	0
42	40	void benchmark_func<__half2, 256, 8u, 7u>(__half2, __half2*) [clone .kd]	0
43	41	void benchmark_func<int, 256, 8u, 7u>(int, int*) [clone .kd]	0
44	42	void benchmark_func<float, 256, 8u, 8u>(float, float*) [clone .kd]	0
45	43	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 8u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
46	44	void benchmark_func<double, 256, 8u, 8u>(double, double*) [clone .kd]	0
47	45	void benchmark_func<__half2, 256, 8u, 8u>(__half2, __half2*) [clone .kd]	0
48	46	void benchmark_func<int, 256, 8u, 8u>(int, int*) [clone .kd]	0
49	47	void benchmark_func<float, 256, 8u, 9u>(float, float*) [clone .kd]	0
50	48	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 9u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
51	49	void benchmark_func<double, 256, 8u, 9u>(double, double*) [clone .kd]	0
52	50	void benchmark_func<__half2, 256, 8u, 9u>(__half2, __half2*) [clone .kd]	0
53	51	void benchmark_func<int, 256, 8u, 9u>(int, int*) [clone .kd]	0
54	52	void benchmark_func<float, 256, 8u, 10u>(float, float*) [clone .kd]	0
55	53	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 10u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
56	54	void benchmark_func<double, 256, 8u, 10u>(double, double*) [clone .kd]	0
57	55	void benchmark_func<__half2, 256, 8u, 10u>(__half2, __half2*) [clone .kd]	0
58	56	void benchmark_func<int, 256, 8u, 10u>(int, int*) [clone .kd]	0
59	57	void benchmark_func<float, 256, 8u, 11u>(float, float*) [clone .kd]	0
60	58	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 11u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
61	59	void benchmark_func<double, 256, 8u, 11u>(double, double*) [clone .kd]	0
62	60	void benchmark_func<__half2, 256, 8u, 11u>(__half2, __half2*) [clone .kd]	0
63	61	void benchmark_func<int, 256, 8u, 11u>(int, int*) [clone .kd]	0
64	62	void benchmark_func<float, 256, 8u, 12u>(float, float*) [clone .kd]	0
65	63	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 12u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
66	64	void benchmark_func<double, 256, 8u, 12u>(double, double*) [clone .kd]	0
67	65	void benchmark_func<__half2, 256, 8u, 12u>(__half2, __half2*) [clone .kd]	0
68	66	void benchmark_func<int, 256, 8u, 12u>(int, int*) [clone .kd]	0
69	67	void benchmark_func<float, 256, 8u, 13u>(float, float*) [clone .kd]	0
70	68	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 13u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
71	69	void benchmark_func<double, 256, 8u, 13u>(double, double*) [clone .kd]	0
72	70	void benchmark_func<__half2, 256, 8u, 13u>(__half2, __half2*) [clone .kd]	0
73	71	void benchmark_func<int, 256, 8u, 13u>(int, int*) [clone .kd]	0
74	72	void benchmark_func<float, 256, 8u, 14u>(float, float*) [clone .kd]	0
75	73	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 14u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
76	74	void benchmark_func<double, 256, 8u, 14u>(double, double*) [clone .kd]	0
77	75	void benchmark_func<__half2, 256, 8u, 14u>(__half2, __half2*) [clone .kd]	0
78	76	void benchmark_func<int, 256, 8u, 14u>(int, int*) [clone .kd]	0
79	77	void benchmark_func<float, 256, 8u, 15u>(float, float*) [clone .kd]	0
80	78	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 15u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
81	79	void benchmark_func<double, 256, 8u, 15u>(double, double*) [clone .kd]	0
82	80	void benchmark_func<__half2, 256, 8u, 15u>(__half2, __half2*) [clone .kd]	0
83	81	void benchmark_func<int, 256, 8u, 15u>(int, int*) [clone .kd]	0
84	82	void benchmark_func<float, 256, 8u, 16u>(float, float*) [clone .kd]	0
85	83	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 16u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
86	84	void benchmark_func<double, 256, 8u, 16u>(double, double*) [clone .kd]	0
87	85	void benchmark_func<__half2, 256, 8u, 16u>(__half2, __half2*) [clone .kd]	0
88	86	void benchmark_func<int, 256, 8u, 16u>(int, int*) [clone .kd]	0
89	87	void benchmark_func<float, 256, 8u, 17u>(float, float*) [clone .kd]	0
90	88	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 17u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
91	89	void benchmark_func<double, 256, 8u, 17u>(double, double*) [clone .kd]	0
92	90	void benchmark_func<__half2, 256, 8u, 17u>(__half2, __half2*) [clone .kd]	0
93	91	void benchmark_func<int, 256, 8u, 17u>(int, int*) [clone .kd]	0
94	92	void benchmark_func<float, 256, 8u, 18u>(float, float*) [clone .kd]	0
95	93	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 18u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
96	94	void benchmark_func<double, 256, 8u, 18u>(double, double*) [clone .kd]	0
97	95	void benchmark_func<__half2, 256, 8u, 18u>(__half2, __half2*) [clone .kd]	0
98	96	void benchmark_func<int, 256, 8u, 18u>(int, int*) [clone .kd]	0
99	97	void benchmark_func<float, 256, 8u, 20u>(float, float*) [clone .kd]	0
100	98	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 20u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
101	99	void benchmark_func<double, 256, 8u, 20u>(double, double*) [clone .kd]	0
102	100	void benchmark_func<__half2, 256, 8u, 20u>(__half2, __half2*) [clone .kd]	0
103	101	void benchmark_func<int, 256, 8u, 20u>(int, int*) [clone .kd]	0
104	102	void benchmark_func<float, 256, 8u, 22u>(float, float*) [clone .kd]	0
105	103	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 22u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
106	104	void benchmark_func<double, 256, 8u, 22u>(double, double*) [clone .kd]	0
107	105	void benchmark_func<__half2, 256, 8u, 22u>(__half2, __half2*) [clone .kd]	0
108	106	void benchmark_func<int, 256, 8u, 22u>(int, int*) [clone .kd]	0
109	107	void benchmark_func<float, 256, 8u, 24u>(float, float*) [clone .kd]	0
110	108	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 24u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
111	109	void benchmark_func<double, 256, 8u, 24u>(double, double*) [clone .kd]	0
112	110	void benchmark_func<__half2, 256, 8u, 24u>(__half2, __half2*) [clone .kd]	0
113	111	void benchmark_func<int, 256, 8u, 24u>(int, int*) [clone .kd]	0
114	112	void benchmark_func<float, 256, 8u, 28u>(float, float*) [clone .kd]	0
115	113	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 28u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
116	114	void benchmark_func<double, 256, 8u, 28u>(double, double*) [clone .kd]	0
117	115	void benchmark_func<__half2, 256, 8u, 28u>(__half2, __half2*) [clone .kd]	0
118	116	void benchmark_func<int, 256, 8u, 28u>(int, int*) [clone .kd]	0
119	117	void benchmark_func<float, 256, 8u, 32u>(float, float*) [clone .kd]	0
120	118	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 32u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
121	119	void benchmark_func<double, 256, 8u, 32u>(double, double*) [clone .kd]	0
122	120	void benchmark_func<__half2, 256, 8u, 32u>(__half2, __half2*) [clone .kd]	0
123	121	void benchmark_func<int, 256, 8u, 32u>(int, int*) [clone .kd]	0
124	122	void benchmark_func<float, 256, 8u, 40u>(float, float*) [clone .kd]	0
125	123	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 40u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
126	124	void benchmark_func<double, 256, 8u, 40u>(double, double*) [clone .kd]	0
127	125	void benchmark_func<__half2, 256, 8u, 40u>(__half2, __half2*) [clone .kd]	0
128	126	void benchmark_func<int, 256, 8u, 40u>(int, int*) [clone .kd]	0
129	127	void benchmark_func<float, 256, 8u, 48u>(float, float*) [clone .kd]	0
130	128	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 48u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
131	129	void benchmark_func<double, 256, 8u, 48u>(double, double*) [clone .kd]	0
132	130	void benchmark_func<__half2, 256, 8u, 48u>(__half2, __half2*) [clone .kd]	0
133	131	void benchmark_func<int, 256, 8u, 48u>(int, int*) [clone .kd]	0
134	132	void benchmark_func<float, 256, 8u, 56u>(float, float*) [clone .kd]	0
135	133	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 56u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
136	134	void benchmark_func<double, 256, 8u, 56u>(double, double*) [clone .kd]	0
137	135	void benchmark_func<__half2, 256, 8u, 56u>(__half2, __half2*) [clone .kd]	0
138	136	void benchmark_func<int, 256, 8u, 56u>(int, int*) [clone .kd]	0
139	137	void benchmark_func<float, 256, 8u, 64u>(float, float*) [clone .kd]	0
140	138	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 64u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
141	139	void benchmark_func<double, 256, 8u, 64u>(double, double*) [clone .kd]	0
142	140	void benchmark_func<__half2, 256, 8u, 64u>(__half2, __half2*) [clone .kd]	0
143	141	void benchmark_func<int, 256, 8u, 64u>(int, int*) [clone .kd]	0
144	142	void benchmark_func<float, 256, 8u, 80u>(float, float*) [clone .kd]	0
145	143	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 80u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
146	144	void benchmark_func<double, 256, 8u, 80u>(double, double*) [clone .kd]	0
147	145	void benchmark_func<__half2, 256, 8u, 80u>(__half2, __half2*) [clone .kd]	0
148	146	void benchmark_func<int, 256, 8u, 80u>(int, int*) [clone .kd]	0
149	147	void benchmark_func<float, 256, 8u, 96u>(float, float*) [clone .kd]	0
150	148	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 96u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
151	149	void benchmark_func<double, 256, 8u, 96u>(double, double*) [clone .kd]	0
152	150	void benchmark_func<__half2, 256, 8u, 96u>(__half2, __half2*) [clone .kd]	0
153	151	void benchmark_func<int, 256, 8u, 96u>(int, int*) [clone .kd]	0
154	152	void benchmark_func<float, 256, 8u, 128u>(float, float*) [clone .kd]	0
155	153	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 128u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
156	154	void benchmark_func<double, 256, 8u, 128u>(double, double*) [clone .kd]	0
157	155	void benchmark_func<__half2, 256, 8u, 128u>(__half2, __half2*) [clone .kd]	0
158	156	void benchmark_func<int, 256, 8u, 128u>(int, int*) [clone .kd]	0
159	157	void benchmark_func<float, 256, 8u, 256u>(float, float*) [clone .kd]	0
160	158	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 256u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
161	159	void benchmark_func<double, 256, 8u, 256u>(double, double*) [clone .kd]	0
162	160	void benchmark_func<__half2, 256, 8u, 256u>(__half2, __half2*) [clone .kd]	0
163	161	void benchmark_func<int, 256, 8u, 256u>(int, int*) [clone .kd]	0
164	162	void benchmark_func<float, 256, 8u, 512u>(float, float*) [clone .kd]	0
165	163	void benchmark_func<HIP_vector_type<float, 2u>, 256, 8u, 512u>(HIP_vector_type<float, 2u>, HIP_vector_type<float, 2u>*) [clone .kd]	0
166	164	void benchmark_func<double, 256, 8u, 512u>(double, double*) [clone .kd]	0
167	165	void benchmark_func<__half2, 256, 8u, 512u>(__half2, __half2*) [clone .kd]	0
168	166	void benchmark_func<int, 256, 8u, 512u>(int, int*) [clone .kd]	0