Hugging Face's logo

Datasets:
pjt222
/
ga104-cuda-kernels

Modalities:
Tabular
Text
Formats:
csv
Languages:
English
Size:
< 1K
Tags:
cuda
sass
gpu-kernels
ampere
ga104
sm_86
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
cubin
stringlengths
36
78
kernel
stringlengths
6
34
total_inst
int64
32
7.02k
HMMA
int64
0
128
IMMA
int64
0
256
LDSM
int64
0
96
LDS
int64
0
480
STS
int64
0
283
LDGSTS
int64
0
48
LDG
int64
0
944
STG
int64
0
224
FFMA
int64
0
512
FADD
int64
0
770
FMUL
int64
0
640
FSEL
int64
0
67
MUFU
int64
0
180
SHFL
int64
0
321
IADD3
int64
0
1.96k
IMAD
int64
2
1.12k
ISETP
int64
1
394
LOP3
int64
0
281
BRA
int64
1
290
EXIT
int64
1
33
BAR
int64
0
8
BARRIER
int64
0
0
WARPSYNC
int64
0
32
NOP
int64
8
78
S2R
int64
1
57
other
int64
3
2.08k
useful_pct
float64
0
50.4
kernels/gemm/sgemm/register_blocked.sm_86.cubin
sgemm_register_blocked
1,016
0
0
0
89
14
0
14
16
512
0
0
0
0
0
45
50
59
15
18
2
2
0
0
10
3
167
50.4
kernels/attention/flash_attention/flash_attn.sm_86.cubin
flash_attn_multihead
1,224
0
0
0
128
14
0
16
2
138
227
135
0
38
160
47
20
18
21
15
1
0
0
0
17
6
221
40.8
kernels/attention/flash_attention/flash_attn_br16_v2_bc128.sm_86.cubin
flash_attn_br16_v2_bc128
1,496
128
0
96
0
46
0
30
32
13
138
326
0
72
9
51
21
23
30
22
3
4
0
6
8
10
428
40.4
kernels/attention/flash_attention/flash_br16_v2_bc128.sm_86.cubin
flash_attn_br16_v2_bc128
1,496
128
0
96
0
46
0
30
32
13
138
326
0
72
9
51
21
23
30
22
3
4
0
6
8
10
428
40.4
kernels/attention/flash_attention/flash_attn.sm_86.cubin
flash_attn_1warp
1,272
0
0
0
128
14
0
16
2
138
227
135
0
38
160
46
16
33
25
26
2
0
0
0
13
3
250
39.3
kernels/elementwise/activations/activations_fast.sm_86.cubin
gelu_kernel
72
0
0
0
0
0
0
4
4
4
4
20
0
4
0
3
3
4
0
1
5
0
0
0
10
2
4
38.9
kernels/elementwise/activations/activations.sm_86.cubin
gelu_kernel
72
0
0
0
0
0
0
4
4
4
4
20
0
4
0
3
3
4
0
1
5
0
0
0
10
2
4
38.9
kernels/attention/flash_attention/flash_wmma.sm_86.cubin
flash_attn_4warp
4,152
0
0
0
382
14
0
16
2
264
770
454
0
70
321
42
16
18
21
21
2
2
0
2
13
6
1,716
35.8
kernels/attention/flash_attention/flash_br16_v2.sm_86.cubin
flash_attn_br16_v2
1,024
64
0
48
0
30
0
30
32
13
74
198
0
40
9
40
55
23
30
22
3
4
0
6
10
10
283
34.1
kernels/attention/cross_attention/cross_attn_v2.sm_86.cubin
cross_attn_v2
1,040
64
0
48
0
30
0
30
32
13
74
198
0
40
9
39
63
23
30
22
3
4
0
6
8
12
292
33.6
kernels/attention/flash_attention/flash_v2_persistent.sm_86.cubin
flash_attn_v2_persistent
1,088
64
0
48
1
31
0
30
32
13
74
198
0
42
9
48
82
33
34
25
1
6
0
6
8
5
298
32.1
kernels/attention/flash_attention/flash_br16_bc128.sm_86.cubin
flash_attn_bc128
4,080
128
0
96
236
283
0
30
32
70
433
640
0
180
161
96
154
59
42
58
17
4
0
5
16
7
1,333
31.2
kernels/attention/flash_attention/flash_v2_persistent_pad.sm_86.cubin
flash_attn_v2_persistent_pad
1,120
64
0
48
1
31
0
30
32
13
74
198
0
42
9
51
97
33
33
25
1
6
0
6
11
5
310
31.2
kernels/attention/flash_attention/flash_br16_regpv_full.sm_86.cubin
flash_attn_br16_regpv_full
1,128
64
0
48
0
30
0
30
32
13
74
198
0
40
9
69
86
23
29
29
3
4
0
6
8
22
311
30.9
kernels/attention/cross_attention/cross_attn_v2_pad.sm_86.cubin
cross_attn_v2_pad
1,184
64
0
48
0
30
0
30
32
13
74
198
0
40
9
43
116
23
29
22
3
4
0
6
8
38
354
29.5
kernels/attention/flash_attention/flash_attn_br16_v2_pad.sm_86.cubin
flash_attn_br16_v2_pad
1,184
64
0
48
0
30
0
30
32
13
74
198
0
40
9
43
116
23
29
22
3
4
0
6
9
38
353
29.5
kernels/attention/flash_attention/flash_br16_v2_pad.sm_86.cubin
flash_attn_br16_v2_pad
1,184
64
0
48
0
30
0
30
32
13
74
198
0
40
9
43
116
23
29
22
3
4
0
6
9
38
353
29.5
kernels/reductions/layernorm/layernorm.sm_86.cubin
layernorm_warp
384
0
0
0
0
0
0
3
1
78
18
13
1
14
14
11
11
14
17
33
3
0
0
0
10
2
141
28.4
kernels/reductions/layernorm/layernorm.sm_86.cubin
layernorm_block
664
0
0
0
5
5
0
12
4
124
42
20
3
21
22
20
17
20
18
55
3
2
0
2
11
2
256
28
kernels/attention/flash_attention/flash_attn_br16_v2_pipeline_pad2.sm_86.cubin
flash_attn_br16_v2_pipeline_pad2
1,256
64
0
48
0
16
24
16
32
13
74
198
0
40
9
130
159
33
39
30
3
4
0
6
11
8
299
27.8
kernels/attention/flash_attention/flash_br16_v2_pipeline_pad.sm_86.cubin
flash_attn_br16_v2_pipeline_pad
1,256
64
0
48
0
16
24
16
32
13
74
198
0
40
9
129
160
33
40
30
3
4
0
6
9
8
300
27.8
kernels/attention/flash_attention/flash_br16_v2_pipeline_pad2.sm_86.cubin
flash_attn_br16_v2_pipeline_pad2
1,256
64
0
48
0
16
24
16
32
13
74
198
0
40
9
130
159
33
39
30
3
4
0
6
11
8
299
27.8
kernels/attention/flash_attention/flash_br16_v2_pipeline.sm_86.cubin
flash_attn_br16_v2_pipeline
1,256
64
0
48
0
16
24
16
32
13
74
198
0
40
9
133
144
33
41
30
3
4
0
6
8
8
312
27.8
kernels/attention/flash_attention/flash_br16_regpv.sm_86.cubin
flash_attn_br16_regpv
3,032
64
0
48
128
110
0
30
32
70
305
384
0
116
161
74
226
51
42
52
17
5
0
7
15
9
1,086
27.1
kernels/attention/flash_attention/flash_br16.sm_86.cubin
flash_attn_br16
3,168
64
0
48
238
179
0
30
32
70
305
416
0
116
161
79
190
56
42
56
17
4
0
5
18
7
1,035
27
kernels/attention/flash_attention/flash_br16_regpv_lean.sm_86.cubin
flash_attn_br16_regpv_lean
3,056
64
0
48
128
110
0
30
32
70
305
384
67
116
194
72
248
71
42
53
17
5
0
8
9
8
975
26.9
kernels/attention/cross_attention/cross_attn.sm_86.cubin
cross_attn_br16
3,088
64
0
48
144
179
0
30
32
55
290
416
0
116
161
102
192
58
42
119
17
4
0
5
16
9
989
26.7
kernels/attention/flash_attention/flash_br16_regpv_lean_qcache.sm_86.cubin
flash_attn_br16_regpv_lean_qcache
3,088
64
0
48
128
110
0
30
32
70
305
384
67
116
194
72
262
71
43
54
17
5
0
10
9
9
988
26.7
kernels/attention/flash_attention/flash_br16_regpv_pad_kv8_w0.sm_86.cubin
flash_attn_br16_regpv_pad
3,120
64
0
48
128
110
0
30
32
70
305
384
0
116
161
93
251
51
42
59
17
5
0
7
8
22
1,117
26.4
kernels/attention/flash_attention/flash_fused.sm_86.cubin
flash_attn_fused
3,256
64
0
48
238
186
0
30
32
70
305
416
0
116
161
122
218
56
42
63
17
4
0
5
19
6
1,038
26.3
kernels/attention/flash_attention/flash_br16_regpv_pad_kv0_w4.sm_86.cubin
flash_attn_br16_regpv_pad
3,136
64
0
48
128
110
0
30
32
70
305
384
0
116
161
98
259
51
42
59
17
5
0
7
8
22
1,120
26.2
kernels/attention/flash_attention/flash_br16_regpv_pad_kv8_w4.sm_86.cubin
flash_attn_br16_regpv_pad
3,136
64
0
48
128
110
0
30
32
70
305
384
0
116
161
94
259
51
42
59
17
5
0
7
8
22
1,124
26.2
kernels/attention/flash_attention/flash_persistent.sm_86.cubin
flash_attn_persistent
3,288
64
0
48
239
206
0
30
32
70
305
416
0
118
161
87
228
71
45
76
1
6
0
5
15
4
1,061
26
kernels/attention/flash_attention/flash_split_q.sm_86.cubin
flash_attn_split_q
3,000
64
0
48
238
205
0
30
64
31
287
384
0
97
161
80
210
47
17
38
2
5
0
5
17
8
962
25.5
kernels/attention/flash_attention/flash_br16_pipeline.sm_86.cubin
flash_attn_br16_pipeline
3,392
64
0
48
238
165
24
16
32
70
305
416
0
116
161
137
267
65
50
64
17
4
0
5
12
5
1,111
25.2
kernels/elementwise/activations/activations_fast.sm_86.cubin
gelu_fast
144
0
0
0
0
0
0
4
4
8
8
12
0
10
0
9
12
12
10
12
5
0
0
0
9
2
27
19.4
kernels/elementwise/activations/activations.sm_86.cubin
gelu_fast
144
0
0
0
0
0
0
4
4
8
8
12
0
10
0
9
12
12
10
12
5
0
0
0
9
2
27
19.4
kernels/gemm/sgemm/tiled.sm_86.cubin
sgemm_tiled
168
0
0
0
64
2
0
2
1
32
0
0
0
0
0
5
8
8
0
3
2
2
0
0
10
4
25
19
kernels/convolution/resblock/resblock.sm_86.cubin
groupnorm_silu_fused
1,200
0
0
0
1
1
0
20
5
129
55
37
0
44
15
127
228
75
59
61
3
0
0
0
16
2
322
18.4
kernels/elementwise/timestep_emb/timestep_emb.sm_86.cubin
timestep_emb_with_silu_scale
72
0
0
0
0
0
0
5
2
2
2
9
0
8
0
1
7
2
0
1
2
0
0
0
15
2
14
18.1
kernels/attention/cross_attention/cross_attn_pipelined.sm_86.cubin
cross_attn_pipelined
4,120
64
0
48
272
181
24
16
112
55
210
400
0
68
161
347
533
121
55
284
3
5
0
6
16
7
1,132
17.7
kernels/reductions/softmax/softmax.sm_86.cubin
softmax_block
248
0
0
0
4
4
0
4
4
9
18
16
4
9
15
12
33
18
12
20
3
4
0
3
12
2
42
17.3
kernels/reductions/groupnorm/groupnorm.sm_86.cubin
groupnorm_nchw
1,000
0
0
0
1
1
0
20
5
112
43
17
0
28
15
116
205
63
41
47
3
0
0
0
10
2
271
17.2
kernels/reductions/softmax/softmax.sm_86.cubin
softmax_warp
128
0
0
0
0
0
0
1
1
9
9
4
1
6
10
4
13
8
11
8
3
0
0
0
8
2
30
17.2
kernels/reductions/groupnorm/groupnorm.sm_86.cubin
groupnorm
1,008
0
0
0
1
1
0
20
5
112
43
17
0
28
15
117
208
64
42
47
3
0
0
0
15
2
268
17.1
kernels/elementwise/activations/activations_fast.sm_86.cubin
silu_kernel
144
0
0
0
0
0
0
4
4
8
8
8
0
10
0
9
14
12
10
12
5
0
0
0
13
2
25
16.7
kernels/elementwise/activations/activations.sm_86.cubin
silu_kernel
144
0
0
0
0
0
0
4
4
8
8
8
0
10
0
9
14
12
10
12
5
0
0
0
13
2
25
16.7
kernels/gemm/igemm/igemm_register_blocked.sm_86.cubin
igemm_register_blocked
1,256
0
64
8
64
14
0
14
64
0
0
129
0
0
0
76
150
59
38
35
2
2
0
16
12
23
486
15.4
kernels/gemm/hgemm/hgemm_16warp_aligned.sm_86.cubin
hgemm_16warp_aligned
440
64
0
32
0
0
8
0
16
0
0
0
0
0
0
35
106
4
15
6
1
4
0
0
9
10
130
14.5
kernels/gemm/sgemm/naive.sm_86.cubin
sgemm_naive
216
0
0
0
0
0
0
58
1
29
0
0
0
0
0
11
42
12
1
11
2
0
0
0
14
4
31
13.4
kernels/convolution/conv2d/conv2d.sm_86.cubin
conv2d_nhwc
2,112
0
0
0
279
7
0
287
1
279
0
0
0
1
0
221
287
134
49
108
2
2
0
0
13
7
435
13.2
kernels/attention/flash_attention/flash_split_q.sm_86.cubin
flash_attn_split_q_reduce
7,016
0
0
0
0
0
0
944
64
375
130
369
0
132
0
1,956
729
230
28
262
33
0
0
0
14
2
1,748
12.5
kernels/elementwise/timestep_emb/timestep_emb.sm_86.cubin
timestep_emb_single
40
0
0
0
0
0
0
0
2
0
0
5
0
4
0
0
2
1
0
1
2
0
0
0
14
1
8
12.5
kernels/elementwise/timestep_emb/timestep_emb.sm_86.cubin
timestep_emb_batch
48
0
0
0
0
0
0
1
2
0
0
5
0
4
0
1
5
2
0
1
2
0
0
0
11
2
12
10.4
kernels/gemm/igemm/igemm_warp_specialized.imma_s02.sm_86.cubin
igemm_warp_specialized
5,656
0
128
16
372
209
48
48
112
288
2
97
0
4
10
459
963
258
173
283
4
8
0
19
47
24
2,084
9.1
kernels/gemm/igemm/igemm_warp_specialized.imma_s04.sm_86.cubin
igemm_warp_specialized
5,656
0
128
16
372
209
48
48
112
288
2
97
0
4
10
459
963
258
173
283
4
8
0
19
47
24
2,084
9.1
kernels/gemm/igemm/igemm_warp_specialized.sm_86.cubin
igemm_warp_specialized
5,656
0
128
16
372
209
48
48
112
288
2
97
0
4
10
459
963
258
173
283
4
8
0
19
47
24
2,084
9.1
kernels/composition/attention_layer/utils.sm_86.cubin
residual_add
48
0
0
0
0
0
0
8
4
0
4
0
0
0
0
3
5
4
0
1
2
0
0
0
12
2
3
8.3
kernels/gemm/igemm/igemm_online_quant_bankfree.sm_86.cubin
igemm_online_quant_bankfree
3,512
0
64
8
220
81
32
32
56
160
2
65
0
4
10
273
654
166
98
177
2
7
0
8
29
14
1,350
8.3
kernels/gemm/igemm/igemm_online_quant_inplace.sm_86.cubin
igemm_online_quant_inplace
3,560
0
64
8
284
129
32
32
56
160
2
65
0
4
10
246
695
166
101
177
2
7
0
8
26
14
1,272
8.2
kernels/gemm/igemm/igemm_online_quant.sm_86.cubin
igemm_online_quant
3,544
0
64
8
284
129
32
32
56
160
2
65
0
4
10
243
678
166
99
177
2
5
0
8
27
14
1,279
8.2
kernels/gemm/hgemm/hgemm_256x128.sm_86.cubin
hgemm_256x128
792
64
0
24
0
48
6
42
32
0
0
0
0
0
0
82
138
67
24
23
2
2
0
8
15
3
212
8.1
kernels/gemm/igemm/igemm_tiled_aggressive.sm_86.cubin
igemm_tiled
608
0
16
4
32
14
0
14
16
0
0
33
0
0
0
61
79
47
25
24
2
2
0
4
12
10
213
8.1
kernels/gemm/igemm/igemm_tiled_handtuned.sm_86.cubin
igemm_tiled
608
0
16
4
32
14
0
14
16
0
0
33
0
0
0
61
79
47
25
24
2
2
0
4
12
10
213
8.1
kernels/gemm/igemm/igemm_tiled.imma_s02.sm_86.cubin
igemm_tiled
608
0
16
4
32
14
0
14
16
0
0
33
0
0
0
61
79
47
25
24
2
2
0
4
12
10
213
8.1
kernels/gemm/igemm/igemm_tiled.imma_s04.sm_86.cubin
igemm_tiled
608
0
16
4
32
14
0
14
16
0
0
33
0
0
0
61
79
47
25
24
2
2
0
4
12
10
213
8.1
kernels/gemm/igemm/igemm_tiled.sm_86.cubin
igemm_tiled
608
0
16
4
32
14
0
14
16
0
0
33
0
0
0
61
79
47
25
24
2
2
0
4
12
10
213
8.1
kernels/gemm/igemm/igemm_8warp_256x256.imma_s02.sm_86.cubin
igemm_8warp_256x256
6,448
0
256
16
480
256
32
96
224
0
0
225
0
0
0
720
1,119
394
281
290
2
2
0
32
78
57
1,888
7.5
kernels/gemm/igemm/igemm_8warp_256x256.imma_s04.sm_86.cubin
igemm_8warp_256x256
6,448
0
256
16
480
256
32
96
224
0
0
225
0
0
0
720
1,119
394
281
290
2
2
0
32
78
57
1,888
7.5
kernels/gemm/igemm/igemm_8warp_256x256.sm_86.cubin
igemm_8warp_256x256
6,448
0
256
16
480
256
32
96
224
0
0
225
0
0
0
720
1,119
394
281
290
2
2
0
32
78
57
1,888
7.5
kernels/gemm/igemm/igemm_persistent.sm_86.cubin
igemm_persistent
3,256
0
128
8
369
161
24
72
112
0
0
113
0
1
0
296
567
268
82
149
1
4
0
17
44
19
821
7.4
kernels/convolution/conv2d/conv2d.sm_86.cubin
conv2d_1x1_nhwc
448
0
0
0
31
7
0
39
1
31
0
0
0
0
0
29
96
39
10
24
2
2
0
0
12
9
116
6.9
kernels/gemm/igemm/igemm_8warp_tribuf.imma_s02.sm_86.cubin
igemm_8warp_tribuf
2,736
0
128
16
312
128
24
72
56
0
0
57
0
0
0
343
493
198
76
116
2
5
0
8
27
15
660
6.8
kernels/gemm/igemm/igemm_8warp_tribuf.imma_s04.sm_86.cubin
igemm_8warp_tribuf
2,736
0
128
16
312
128
24
72
56
0
0
57
0
0
0
343
493
198
76
116
2
5
0
8
27
15
660
6.8
kernels/gemm/igemm/igemm_8warp_tribuf.sm_86.cubin
igemm_8warp_tribuf
2,736
0
128
16
312
128
24
72
56
0
0
57
0
0
0
343
493
198
76
116
2
5
0
8
27
15
660
6.8
kernels/gemm/hgemm/hgemm_tiled_direct.sm_86.cubin
hgemm_tiled_direct
992
64
0
24
0
48
16
48
32
0
0
0
0
0
0
80
189
99
25
43
2
2
0
8
10
6
296
6.5
kernels/gemm/igemm/igemm_8warp_256.sm_86.cubin
igemm_8warp_256
3,792
0
128
8
368
160
24
72
112
0
0
113
0
0
0
475
696
253
137
163
2
2
0
16
40
21
1,002
6.4
kernels/gemm/igemm/igemm_sparse_tiled_ldsm_fused.sm_86.cubin
igemm_sparse_tiled
1,064
0
32
8
40
64
2
64
32
0
0
33
0
0
0
98
187
126
58
26
3
2
0
0
13
9
267
6.1
kernels/gemm/igemm/igemm_sparse_tiled.sm_86.cubin
igemm_sparse_tiled
1,064
0
32
8
40
64
2
64
32
0
0
33
0
0
0
98
187
126
58
26
3
2
0
0
13
9
267
6.1
kernels/gemm/hgemm/hgemm_16warp.sm_86.cubin
hgemm_16warp
544
32
0
16
0
32
4
28
16
0
0
0
0
0
0
50
138
45
15
15
2
2
0
4
11
7
127
5.9
experiments/rust-experiments/cymatic_oxide/gather_nvcc.sm_86.cubin
gather_sum
120
0
0
0
0
0
0
14
0
0
7
0
0
1
0
10
29
12
3
8
2
0
0
0
13
2
19
5.8
kernels/gemm/igemm/igemm_8warp.imma_s02.sm_86.cubin
igemm_8warp
2,080
0
64
8
184
96
16
48
56
0
0
57
0
0
0
269
392
145
76
90
2
2
0
8
27
13
527
5.8
kernels/gemm/igemm/igemm_8warp.imma_s04.sm_86.cubin
igemm_8warp
2,080
0
64
8
184
96
16
48
56
0
0
57
0
0
0
269
392
145
76
90
2
2
0
8
27
13
527
5.8
kernels/gemm/igemm/igemm_8warp.sm_86.cubin
igemm_8warp
2,080
0
64
8
184
96
16
48
56
0
0
57
0
0
0
269
392
145
76
90
2
2
0
8
27
13
527
5.8
kernels/gemm/igemm/igemm_sparse_tiled_ldsm.sm_86.cubin
igemm_sparse_tiled
1,120
0
32
8
52
74
4
62
32
0
0
33
0
0
0
131
176
121
32
40
3
4
0
0
10
9
297
5.8
kernels/gemm/igemm/igemm_pipelined_cpasync.imma_s02.sm_86.cubin
igemm_pipelined_cpasync
1,136
0
32
8
64
64
16
48
16
0
0
33
0
0
0
170
140
104
23
39
2
2
0
4
15
8
348
5.7
kernels/gemm/igemm/igemm_pipelined_cpasync.imma_s04.sm_86.cubin
igemm_pipelined_cpasync
1,136
0
32
8
64
64
16
48
16
0
0
33
0
0
0
170
140
104
23
39
2
2
0
4
15
8
348
5.7
kernels/gemm/igemm/igemm_pipelined_cpasync.sm_86.cubin
igemm_pipelined_cpasync
1,136
0
32
8
64
64
16
48
16
0
0
33
0
0
0
170
140
104
23
39
2
2
0
4
15
8
348
5.7
kernels/gemm/hgemm/hgemm_16warp_persistent.sm_86.cubin
hgemm_16warp_persistent
584
32
0
16
0
32
4
28
16
0
0
0
0
1
0
55
137
51
17
16
2
3
0
5
12
5
152
5.5
kernels/gemm/igemm/igemm_sparse_tiled_persistent.sm_86.cubin
igemm_sparse_tiled_persistent
1,184
0
32
8
137
63
4
62
32
0
0
33
0
1
0
102
167
113
10
32
1
3
0
1
9
6
368
5.5
kernels/gemm/igemm/igemm_pipelined_cpasync_perchannel.sm_86.cubin
igemm_pipelined_cpasync_perchannel
1,624
0
32
8
92
80
16
88
28
0
28
28
0
0
0
235
292
122
48
59
2
2
0
4
18
15
427
5.4
kernels/gemm/igemm/igemm_pipelined.imma_s02.sm_86.cubin
igemm_pipelined
1,256
0
32
8
64
64
0
64
16
0
0
33
0
0
0
184
212
134
44
7
2
3
0
4
8
6
371
5.2
kernels/gemm/igemm/igemm_pipelined.imma_s04.sm_86.cubin
igemm_pipelined
1,256
0
32
8
64
64
0
64
16
0
0
33
0
0
0
184
212
134
44
7
2
3
0
4
8
6
371
5.2
kernels/gemm/igemm/igemm_pipelined.sm_86.cubin
igemm_pipelined
1,256
0
32
8
64
64
0
64
16
0
0
33
0
0
0
184
212
134
44
7
2
3
0
4
8
6
371
5.2
kernels/gemm/igemm/igemm_pipelined_cpasync_bk64.sm_86.cubin
igemm_pipelined_cpasync_bk64
1,960
0
64
16
128
128
32
96
16
0
0
33
0
0
0
331
231
214
23
71
2
2
0
4
14
7
548
4.9
kernels/gemm/hgemm_sparse/hgemm_sparse_tiled.sm_86.cubin
hgemm_sparse_tiled
720
32
0
40
0
32
4
36
32
0
0
0
0
0
0
69
93
74
12
31
3
2
0
0
10
4
246
4.4
kernels/gemm/igemm/igemm.sm_86.cubin
igemm_wmma
432
0
10
0
0
0
0
50
8
0
0
9
0
0
0
52
127
24
15
13
3
0
0
17
15
16
73
4.4
experiments/rust-experiments/cymatic_oxide/gather_oxide.sm_86.cubin
gather_sum
80
0
0
0
1
0
0
0
0
0
3
0
0
0
0
2
16
11
0
7
5
0
0
0
12
2
21
3.8
kernels/gemm/hgemm/hgemm_tiled.sm_86.cubin
hgemm_tiled
1,728
64
0
24
56
80
16
48
56
0
0
0
0
0
0
221
374
131
74
91
2
2
0
8
28
12
441
3.7
kernels/gemm/hgemm/hgemm_16warp_epi.sm_86.cubin
hgemm_16warp_epi
952
32
0
16
28
40
8
24
28
0
0
0
0
0
0
120
205
65
38
47
2
2
0
4
20
18
255
3.4
End of preview. Expand in Data Studio

GA104 Hand-Optimized CUDA Kernel Corpus

A measurement corpus of hand-optimized CUDA / SASS kernels targeting the RTX 3070 Ti (GA104, sm_86, Ampere). Every kernel is written without cuBLAS, cuDNN, or PyTorch in the optimized path; vendor libraries are linked only for measured comparison under kernels/reference/. This dataset is for SASS and GPU-optimization researchers — it pairs each .cu source with its compiled machine code and its disassembly, so the exact instruction stream a kernel produced on this toolchain can be studied without owning the hardware.

The four laws of GA104

The corpus is organized around four empirically derived constraints. The observations behind each are in docs/gpu_reflections.md.

  1. Feed Tensor Cores continuously. Overlap loads with HMMA / IMMA. At ≥8 warps, cp.async benefit depends on the compute/load ratio — helpful when compute is short, harmful when compute is long.
  2. Read each byte of DRAM at most once per kernel. im2col converts 9× re-reads to 1×; implicit GEMM eliminates the column buffer.
  3. Fill the warp schedulers. 32 warps/SM is ideal, 8 sufficient; below 8 indicates a structural problem.
  4. Never cross the 50 KB shared-memory cliff per block. Blocks at

    50 KB drop to 1 block/SM (4 warps), a measured 2× regression.

sm_86 only

Every artifact in this corpus is compiled for compute capability 8.6 (sm_86) only. The cubins are not portable to other architectures and will not load on a non-Ampere or non-GA10x device. This is deliberate: the corpus is a single-target measurement record, not a portable kernel library. There are no fat binaries and no PTX-only fallbacks.

Provenance

This corpus was built and published by scripts/publish_hf.R in the source repository. The stamp below is filled in at publish time.

  • Source commit: cd06d0a628464a4186f702ac8ca7ee3fd629d8b9
  • Build date: 2026-05-21 13:49:49 CEST
  • Compiler: CUDA 13.2, nvcc V13.2.78
  • Cubin build line: nvcc -arch=sm_86 -O2 --cubin
  • Disassembly: cuobjdump -sass (via scripts/build.R disasm)
  • Measurement hardware: RTX 3070 Ti Laptop (GA104, 46-SM bin)

What the .cubin and .sass files are

*.sm_86.cubin files are GPU machine code — ELF containers of sm_86 SASS, loaded through the CUDA Driver API (cuModuleLoad). They are not host executables; running one on a CPU does nothing. The matching *.sm_86.sass files are cuobjdump -sass disassembly of those cubins, provided so the instruction stream is greppable without a CUDA install.

The repository's local R package cuasmR reads a cubin via nvdisasm, indexes instructions by .text byte offset, and patches at the byte level — a byte-identical roundtrip (compile → disasm → reassemble) is part of its test surface. See docs/cuasm_r.md.

Headline performance

Measured on RTX 3070 Ti Laptop (GA104, sm_86, 46-SM bin), CUDA 13.2 / nvcc V13.2.78, driver 595.97. Sparse 2:4 figures are dense-equivalent (the multiply count the sparse pattern would do as dense work).

Kernel Size GFLOPS / TOPS % peak
Sparse HGEMM 2:4 2048³ 41,721 (eq) 24.0%
Sparse INT8 mma.sp 2048³ 39,674 (eq) 11.4%
HGEMM 16-warp 4096³ 31,910 18.3%
IGEMM 128×256 4096³ 27,591 7.9%
Flash Attention v2 seq=1024 b=8 h=8 11,453 6.6%
Conv2d implicit GEMM 64×64, 320ch 6,687 3.8%

The headline figure is 41,721 GFLOPS dense-equivalent for sparse 2:4 HGEMM at 2048³. Full per-kernel tables and the phase progression (naive SGEMM → sparse INT8, 90.5× cumulative) are in docs/inventory.md.

File layout

The corpus distinguishes tracked sources from the generated supplement rebuilt at publish time.

Path Contents
kernels/ Tracked .cu / .cuh kernel sources, grouped by family
kernels/**/*_handtuned.sm_86.cubin Tracked hand-patched cubins (SASS hand-edits)
generated/ Rebuilt .sm_86.cubin + .sm_86.sass set, build output (kernels of record only — test/probe binaries and negative-result experiment variants excluded)
data/ Regenerable CSV/JSON: baselines, register audit, SASS histogram
docs/ Researcher-facing analyses and references
AGENTS.md Hardware constants, toolchain, conventions, the four laws
SHA256SUMS SHA-256 of every cubin in the corpus
LICENSE MIT

generated/ is build output — it is rebuilt deterministically from the tracked kernels/ sources with the build line above. Treat kernels/ as canonical and generated/ as a convenience supplement.

Reproduction

The single source of truth and the only supported reproduction path is the GitHub repository:

https://github.com/pjt222/bare-metal

From a fresh clone on an sm_86 host: make reproduce (setup → verify → build → bench). This dataset is a published snapshot; it is not the build system.

License

MIT. See LICENSE. Vendor libraries linked under kernels/reference/ for comparison are not redistributed here — only the project's own benchmark wrappers are included.
Downloads last month
294
Number of rows:
123
Total file size:
36.9 MB