- Reland D75563906 #4865
Open
- Reland D75563906 #4865
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
✅ Deploy Preview for pytorch-fbgemm-docs ready!
To edit notification comments on pull requests, go to your Netlify project configuration. |
|
This pull request was exported from Phabricator. Differential Revision: D77698275 |
flaviotruzzi
force-pushed
the
export-D77698275
branch
from
a1d116e to
2fb84a9
Compare
flaviotruzzi
added a commit
to flaviotruzzi/FBGEMM
that referenced
this pull request
Sep 15, 2025
Summary:
Reland D75563906 that was backed out, with fixes.
The problem was the grid was not being big enough given the config.
Further ensured vectorization, which allows 1.4Tb/s.
Test:
```
Running correctness tests...
Testing correctness for dtype: torch.float32
✓ Passed: shape (32, 32)
✓ Passed: shape (64, 32)
✓ Passed: shape (256, 128)
✓ Passed: shape (512, 1024)
✓ Passed: shape (1024, 2048)
✓ Passed: shape (2048, 2048)
✓ Passed: shape (4096, 16384)
✓ Passed: shape (70000, 64)
✓ Passed: shape (131072, 512)
✓ Passed: shape (1000, 520)
✓ Passed: shape (4005, 4005)
✓ Passed: shape (10000, 1000)
✓ Passed: shape (1024, 10000)
✓ Passed: shape (8192, 4096)
✓ Passed: shape (10000, 10000)
✓ Passed: shape (3072, 10000)
✓ Passed: shape (6144, 10000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (512, 1536)
✓ Passed: shape (512, 6144)
✓ Passed: shape (512, 10240)
✓ Passed: shape (1000, 1000)
✓ Passed: shape (2000, 2000)
✓ Passed: shape (10240, 10240)
✓ Passed: shape (384, 128)
✓ Passed: shape (2048, 1024)
✓ Passed: shape (267, 513)
✓ Passed: shape (67, 123479)
✓ Passed: shape (1024, 123479)
✓ Passed: shape (1234154, 512)
✓ Passed: shape (2048, 66679)
✓ Passed: shape (200, 256)
✓ Passed: shape (1000, 256)
✓ Passed: shape (6000, 256)
✓ Passed: shape (6272, 256)
✓ Passed: shape (200, 512)
✓ Passed: shape (1000, 512)
✓ Passed: shape (6000, 512)
✓ Passed: shape (6272, 512)
✓ Passed: shape (200, 1024)
✓ Passed: shape (1000, 1024)
✓ Passed: shape (6000, 1024)
✓ Passed: shape (6272, 1024)
✓ Passed: shape (200, 2048)
✓ Passed: shape (1000, 2048)
✓ Passed: shape (6000, 2048)
✓ Passed: shape (6272, 2048)
✓ Passed: shape (200, 3072)
✓ Passed: shape (1000, 3072)
✓ Passed: shape (6000, 3072)
✓ Passed: shape (6272, 3072)
✓ Passed: shape (3000000, 512)
Testing correctness for dtype: torch.float16
✓ Passed: shape (32, 32)
✓ Passed: shape (64, 32)
✓ Passed: shape (256, 128)
✓ Passed: shape (512, 1024)
✓ Passed: shape (1024, 2048)
✓ Passed: shape (2048, 2048)
✓ Passed: shape (4096, 16384)
✓ Passed: shape (70000, 64)
✓ Passed: shape (131072, 512)
✓ Passed: shape (1000, 520)
✓ Passed: shape (4005, 4005)
✓ Passed: shape (10000, 1000)
✓ Passed: shape (1024, 10000)
✓ Passed: shape (8192, 4096)
✓ Passed: shape (10000, 10000)
✓ Passed: shape (3072, 10000)
✓ Passed: shape (6144, 10000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (512, 1536)
✓ Passed: shape (512, 6144)
✓ Passed: shape (512, 10240)
✓ Passed: shape (1000, 1000)
✓ Passed: shape (2000, 2000)
✓ Passed: shape (10240, 10240)
✓ Passed: shape (384, 128)
✓ Passed: shape (2048, 1024)
✓ Passed: shape (267, 513)
✓ Passed: shape (67, 123479)
✓ Passed: shape (1024, 123479)
✓ Passed: shape (1234154, 512)
✓ Passed: shape (2048, 66679)
✓ Passed: shape (200, 256)
✓ Passed: shape (1000, 256)
✓ Passed: shape (6000, 256)
✓ Passed: shape (6272, 256)
✓ Passed: shape (200, 512)
✓ Passed: shape (1000, 512)
✓ Passed: shape (6000, 512)
✓ Passed: shape (6272, 512)
✓ Passed: shape (200, 1024)
✓ Passed: shape (1000, 1024)
✓ Passed: shape (6000, 1024)
✓ Passed: shape (6272, 1024)
✓ Passed: shape (200, 2048)
✓ Passed: shape (1000, 2048)
✓ Passed: shape (6000, 2048)
✓ Passed: shape (6272, 2048)
✓ Passed: shape (200, 3072)
✓ Passed: shape (1000, 3072)
✓ Passed: shape (6000, 3072)
✓ Passed: shape (6272, 3072)
✓ Passed: shape (3000000, 512)
Testing correctness for dtype: torch.bfloat16
✓ Passed: shape (32, 32)
✓ Passed: shape (64, 32)
✓ Passed: shape (256, 128)
✓ Passed: shape (512, 1024)
✓ Passed: shape (1024, 2048)
✓ Passed: shape (2048, 2048)
✓ Passed: shape (4096, 16384)
✓ Passed: shape (70000, 64)
✓ Passed: shape (131072, 512)
✓ Passed: shape (1000, 520)
✓ Passed: shape (4005, 4005)
✓ Passed: shape (10000, 1000)
✓ Passed: shape (1024, 10000)
✓ Passed: shape (8192, 4096)
✓ Passed: shape (10000, 10000)
✓ Passed: shape (3072, 10000)
✓ Passed: shape (6144, 10000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (512, 1536)
✓ Passed: shape (512, 6144)
✓ Passed: shape (512, 10240)
✓ Passed: shape (1000, 1000)
✓ Passed: shape (2000, 2000)
✓ Passed: shape (10240, 10240)
✓ Passed: shape (384, 128)
✓ Passed: shape (2048, 1024)
✓ Passed: shape (267, 513)
✓ Passed: shape (67, 123479)
✓ Passed: shape (1024, 123479)
✓ Passed: shape (1234154, 512)
✓ Passed: shape (2048, 66679)
✓ Passed: shape (200, 256)
✓ Passed: shape (1000, 256)
✓ Passed: shape (6000, 256)
✓ Passed: shape (6272, 256)
✓ Passed: shape (200, 512)
✓ Passed: shape (1000, 512)
✓ Passed: shape (6000, 512)
✓ Passed: shape (6272, 512)
✓ Passed: shape (200, 1024)
✓ Passed: shape (1000, 1024)
✓ Passed: shape (6000, 1024)
✓ Passed: shape (6272, 1024)
✓ Passed: shape (200, 2048)
✓ Passed: shape (1000, 2048)
✓ Passed: shape (6000, 2048)
✓ Passed: shape (6272, 2048)
✓ Passed: shape (200, 3072)
✓ Passed: shape (1000, 3072)
✓ Passed: shape (6000, 3072)
✓ Passed: shape (6272, 3072)
✓ Passed: shape (3000000, 512)
All correctness tests passed!
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:31<00:00, 1.68it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:27<00:00, 1.90it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:28<00:00, 1.86it/s]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [01:27<00:00, 29.29s/it]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:31<00:00, 1.69it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:31<00:00, 1.70it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:30<00:00, 1.73it/s]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [01:33<00:00, 31.02s/it]
[------------------------------------ Not vectorized ------------------------------------]
| fwd, torch.float32 | fwd, torch.float16 | fwd, torch.bfloat16
1 threads: -------------------------------------------------------------------------------
32, 32 | 29.2 | 20.6 | 20.3
64, 32 | 29.3 | 20.5 | 21.2
256, 128 | 23.5 | 20.4 | 20.0
512, 1024 | 22.6 | 20.6 | 20.0
1024, 2048 | 52.9 | 40.9 | 40.9
2048, 2048 | 88.8 | 58.1 | 58.1
4096, 16384 | 1327.5 | 918.2 | 921.5
70000, 64 | 94.0 | 67.1 | 67.1
131072, 512 | 908.2 | 710.9 | 716.4
1000, 520 | 22.4 | 20.1 | 20.6
4005, 4005 | 339.3 | 237.7 | 238.7
10000, 1000 | 207.8 | 129.9 | 130.7
1024, 10000 | 357.8 | 272.0 | 277.6
8192, 4096 | 669.1 | 463.8 | 465.1
10000, 10000 | 1934.2 | 1283.3 | 1287.1
3072, 10000 | 676.7 | 484.2 | 484.9
6144, 10000 | 1165.3 | 767.3 | 770.7
1024, 20000 | 703.6 | 577.5 | 578.6
512, 1536 | 30.6 | 25.8 | 25.6
512, 6144 | 99.0 | 80.6 | 80.6
512, 10240 | 205.4 | 128.6 | 128.5
1000, 1000 | 30.2 | 24.0 | 23.8
2000, 2000 | 81.0 | 56.4 | 56.7
10240, 10240 | 1963.8 | 1323.1 | 1326.9
384, 128 | 20.7 | 20.5 | 20.0
2048, 1024 | 45.3 | 33.0 | 33.0
267, 513 | 20.7 | 20.1 | 19.8
67, 123479 | 2322.7 | 1230.8 | 1313.2
1024, 123479 | 4244.2 | 3485.9 | 3491.3
1234154, 512 | 6838.8 | 5890.0 | 5956.4
2048, 66679 | 3304.3 | 2477.3 | 2483.4
200, 256 | 20.5 | 19.7 | 19.7
1000, 256 | 20.5 | 19.9 | 19.6
6000, 256 | 32.9 | 23.7 | 23.8
6272, 256 | 34.0 | 24.8 | 24.9
200, 512 | 21.0 | 19.7 | 20.2
1000, 512 | 20.3 | 20.2 | 19.8
6000, 512 | 56.4 | 38.4 | 38.6
6272, 512 | 59.0 | 40.3 | 40.9
200, 1024 | 20.7 | 19.7 | 20.0
1000, 1024 | 30.2 | 24.1 | 24.0
6000, 1024 | 118.6 | 67.4 | 67.9
6272, 1024 | 123.8 | 73.3 | 72.6
200, 2048 | 28.9 | 26.3 | 26.2
1000, 2048 | 52.3 | 40.5 | 40.2
6000, 2048 | 238.6 | 159.3 | 159.2
6272, 2048 | 246.8 | 165.6 | 165.7
200, 3072 | 39.9 | 36.3 | 36.2
1000, 3072 | 75.0 | 56.8 | 56.5
6000, 3072 | 352.0 | 240.9 | 241.9
6272, 3072 | 365.4 | 249.1 | 250.1
3000000, 512 | 16525.9 | 14259.5 | 14365.3
Times are in microseconds (us).
[-------------------------------------- Vectorized --------------------------------------]
| fwd, torch.float32 | fwd, torch.float16 | fwd, torch.bfloat16
1 threads: -------------------------------------------------------------------------------
32, 32 | 19.4 | 19.6 | 19.6
64, 32 | 19.6 | 19.5 | 20.1
256, 128 | 19.4 | 19.6 | 20.2
512, 1024 | 19.3 | 19.9 | 20.0
1024, 2048 | 30.2 | 29.4 | 29.3
2048, 2048 | 42.4 | 35.1 | 35.0
4096, 16384 | 613.5 | 562.1 | 564.2
70000, 64 | 50.2 | 50.1 | 50.2
131072, 512 | 548.1 | 467.3 | 470.2
1000, 520 | 19.7 | 20.3 | 19.8
4005, 4005 | 206.0 | 240.6 | 240.6
10000, 1000 | 98.3 | 75.4 | 75.1
1024, 10000 | 250.9 | 200.9 | 208.1
8192, 4096 | 317.8 | 285.0 | 286.0
10000, 10000 | 857.6 | 742.6 | 743.9
3072, 10000 | 347.1 | 320.1 | 320.0
6144, 10000 | 501.9 | 433.7 | 435.7
1024, 20000 | 486.9 | 471.7 | 471.3
512, 1536 | 21.9 | 21.7 | 21.6
512, 6144 | 64.8 | 62.7 | 62.9
512, 10240 | 139.6 | 100.2 | 100.2
1000, 1000 | 19.3 | 19.8 | 20.5
2000, 2000 | 36.5 | 34.3 | 34.3
10240, 10240 | 849.1 | 765.9 | 767.3
384, 128 | 20.0 | 20.0 | 20.1
2048, 1024 | 22.4 | 21.5 | 21.6
267, 513 | 20.6 | 20.2 | 19.7
67, 123479 | 2354.0 | 998.4 | 996.2
1024, 123479 | 3134.1 | 3418.3 | 3427.5
1234154, 512 | 5083.0 | 4251.9 | 4278.5
2048, 66679 | 2052.4 | 2323.9 | 2325.0
200, 256 | 19.7 | 19.7 | 19.9
1000, 256 | 20.0 | 19.5 | 19.6
6000, 256 | 20.2 | 19.4 | 19.7
6272, 256 | 19.9 | 19.7 | 20.2
200, 512 | 19.7 | 19.9 | 19.5
1000, 512 | 20.1 | 19.8 | 19.6
6000, 512 | 23.1 | 20.9 | 21.3
6272, 512 | 24.0 | 22.4 | 22.9
200, 1024 | 19.4 | 19.8 | 19.7
1000, 1024 | 19.7 | 19.5 | 20.1
6000, 1024 | 51.9 | 34.1 | 34.6
6272, 1024 | 55.5 | 39.0 | 37.9
200, 2048 | 24.5 | 24.4 | 24.4
1000, 2048 | 30.3 | 29.1 | 28.9
6000, 2048 | 109.2 | 96.5 | 96.9
6272, 2048 | 111.4 | 100.1 | 100.3
200, 3072 | 33.0 | 33.5 | 33.4
1000, 3072 | 41.4 | 39.5 | 39.5
6000, 3072 | 156.3 | 146.5 | 146.9
6272, 3072 | 160.9 | 150.2 | 151.0
3000000, 512 | 12385.8 | 10336.7 | 10366.0
Times are in microseconds (us).
```
Reviewed By: q10
Differential Revision: D77698275
|
@flaviotruzzi has exported this pull request. If you are a Meta employee, you can view the originating diff in D77698275. |
Summary:
Reland D75563906 that was backed out, with fixes.
The problem was the grid was not being big enough given the config.
Further ensured vectorization, which allows 1.4Tb/s.
Test:
```
Running correctness tests...
Testing correctness for dtype: torch.float32
✓ Passed: shape (32, 32)
✓ Passed: shape (64, 32)
✓ Passed: shape (256, 128)
✓ Passed: shape (512, 1024)
✓ Passed: shape (1024, 2048)
✓ Passed: shape (2048, 2048)
✓ Passed: shape (4096, 16384)
✓ Passed: shape (70000, 64)
✓ Passed: shape (131072, 512)
✓ Passed: shape (1000, 520)
✓ Passed: shape (4005, 4005)
✓ Passed: shape (10000, 1000)
✓ Passed: shape (1024, 10000)
✓ Passed: shape (8192, 4096)
✓ Passed: shape (10000, 10000)
✓ Passed: shape (3072, 10000)
✓ Passed: shape (6144, 10000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (512, 1536)
✓ Passed: shape (512, 6144)
✓ Passed: shape (512, 10240)
✓ Passed: shape (1000, 1000)
✓ Passed: shape (2000, 2000)
✓ Passed: shape (10240, 10240)
✓ Passed: shape (384, 128)
✓ Passed: shape (2048, 1024)
✓ Passed: shape (267, 513)
✓ Passed: shape (67, 123479)
✓ Passed: shape (1024, 123479)
✓ Passed: shape (1234154, 512)
✓ Passed: shape (2048, 66679)
✓ Passed: shape (200, 256)
✓ Passed: shape (1000, 256)
✓ Passed: shape (6000, 256)
✓ Passed: shape (6272, 256)
✓ Passed: shape (200, 512)
✓ Passed: shape (1000, 512)
✓ Passed: shape (6000, 512)
✓ Passed: shape (6272, 512)
✓ Passed: shape (200, 1024)
✓ Passed: shape (1000, 1024)
✓ Passed: shape (6000, 1024)
✓ Passed: shape (6272, 1024)
✓ Passed: shape (200, 2048)
✓ Passed: shape (1000, 2048)
✓ Passed: shape (6000, 2048)
✓ Passed: shape (6272, 2048)
✓ Passed: shape (200, 3072)
✓ Passed: shape (1000, 3072)
✓ Passed: shape (6000, 3072)
✓ Passed: shape (6272, 3072)
✓ Passed: shape (3000000, 512)
Testing correctness for dtype: torch.float16
✓ Passed: shape (32, 32)
✓ Passed: shape (64, 32)
✓ Passed: shape (256, 128)
✓ Passed: shape (512, 1024)
✓ Passed: shape (1024, 2048)
✓ Passed: shape (2048, 2048)
✓ Passed: shape (4096, 16384)
✓ Passed: shape (70000, 64)
✓ Passed: shape (131072, 512)
✓ Passed: shape (1000, 520)
✓ Passed: shape (4005, 4005)
✓ Passed: shape (10000, 1000)
✓ Passed: shape (1024, 10000)
✓ Passed: shape (8192, 4096)
✓ Passed: shape (10000, 10000)
✓ Passed: shape (3072, 10000)
✓ Passed: shape (6144, 10000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (512, 1536)
✓ Passed: shape (512, 6144)
✓ Passed: shape (512, 10240)
✓ Passed: shape (1000, 1000)
✓ Passed: shape (2000, 2000)
✓ Passed: shape (10240, 10240)
✓ Passed: shape (384, 128)
✓ Passed: shape (2048, 1024)
✓ Passed: shape (267, 513)
✓ Passed: shape (67, 123479)
✓ Passed: shape (1024, 123479)
✓ Passed: shape (1234154, 512)
✓ Passed: shape (2048, 66679)
✓ Passed: shape (200, 256)
✓ Passed: shape (1000, 256)
✓ Passed: shape (6000, 256)
✓ Passed: shape (6272, 256)
✓ Passed: shape (200, 512)
✓ Passed: shape (1000, 512)
✓ Passed: shape (6000, 512)
✓ Passed: shape (6272, 512)
✓ Passed: shape (200, 1024)
✓ Passed: shape (1000, 1024)
✓ Passed: shape (6000, 1024)
✓ Passed: shape (6272, 1024)
✓ Passed: shape (200, 2048)
✓ Passed: shape (1000, 2048)
✓ Passed: shape (6000, 2048)
✓ Passed: shape (6272, 2048)
✓ Passed: shape (200, 3072)
✓ Passed: shape (1000, 3072)
✓ Passed: shape (6000, 3072)
✓ Passed: shape (6272, 3072)
✓ Passed: shape (3000000, 512)
Testing correctness for dtype: torch.bfloat16
✓ Passed: shape (32, 32)
✓ Passed: shape (64, 32)
✓ Passed: shape (256, 128)
✓ Passed: shape (512, 1024)
✓ Passed: shape (1024, 2048)
✓ Passed: shape (2048, 2048)
✓ Passed: shape (4096, 16384)
✓ Passed: shape (70000, 64)
✓ Passed: shape (131072, 512)
✓ Passed: shape (1000, 520)
✓ Passed: shape (4005, 4005)
✓ Passed: shape (10000, 1000)
✓ Passed: shape (1024, 10000)
✓ Passed: shape (8192, 4096)
✓ Passed: shape (10000, 10000)
✓ Passed: shape (3072, 10000)
✓ Passed: shape (6144, 10000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (1024, 20000)
✓ Passed: shape (512, 1536)
✓ Passed: shape (512, 6144)
✓ Passed: shape (512, 10240)
✓ Passed: shape (1000, 1000)
✓ Passed: shape (2000, 2000)
✓ Passed: shape (10240, 10240)
✓ Passed: shape (384, 128)
✓ Passed: shape (2048, 1024)
✓ Passed: shape (267, 513)
✓ Passed: shape (67, 123479)
✓ Passed: shape (1024, 123479)
✓ Passed: shape (1234154, 512)
✓ Passed: shape (2048, 66679)
✓ Passed: shape (200, 256)
✓ Passed: shape (1000, 256)
✓ Passed: shape (6000, 256)
✓ Passed: shape (6272, 256)
✓ Passed: shape (200, 512)
✓ Passed: shape (1000, 512)
✓ Passed: shape (6000, 512)
✓ Passed: shape (6272, 512)
✓ Passed: shape (200, 1024)
✓ Passed: shape (1000, 1024)
✓ Passed: shape (6000, 1024)
✓ Passed: shape (6272, 1024)
✓ Passed: shape (200, 2048)
✓ Passed: shape (1000, 2048)
✓ Passed: shape (6000, 2048)
✓ Passed: shape (6272, 2048)
✓ Passed: shape (200, 3072)
✓ Passed: shape (1000, 3072)
✓ Passed: shape (6000, 3072)
✓ Passed: shape (6272, 3072)
✓ Passed: shape (3000000, 512)
All correctness tests passed!
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:31<00:00, 1.68it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:27<00:00, 1.90it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:28<00:00, 1.86it/s]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [01:27<00:00, 29.29s/it]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:31<00:00, 1.69it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:31<00:00, 1.70it/s]
100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 53/53 [00:30<00:00, 1.73it/s]
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 3/3 [01:33<00:00, 31.02s/it]
[------------------------------------ Not vectorized ------------------------------------]
| fwd, torch.float32 | fwd, torch.float16 | fwd, torch.bfloat16
1 threads: -------------------------------------------------------------------------------
32, 32 | 29.2 | 20.6 | 20.3
64, 32 | 29.3 | 20.5 | 21.2
256, 128 | 23.5 | 20.4 | 20.0
512, 1024 | 22.6 | 20.6 | 20.0
1024, 2048 | 52.9 | 40.9 | 40.9
2048, 2048 | 88.8 | 58.1 | 58.1
4096, 16384 | 1327.5 | 918.2 | 921.5
70000, 64 | 94.0 | 67.1 | 67.1
131072, 512 | 908.2 | 710.9 | 716.4
1000, 520 | 22.4 | 20.1 | 20.6
4005, 4005 | 339.3 | 237.7 | 238.7
10000, 1000 | 207.8 | 129.9 | 130.7
1024, 10000 | 357.8 | 272.0 | 277.6
8192, 4096 | 669.1 | 463.8 | 465.1
10000, 10000 | 1934.2 | 1283.3 | 1287.1
3072, 10000 | 676.7 | 484.2 | 484.9
6144, 10000 | 1165.3 | 767.3 | 770.7
1024, 20000 | 703.6 | 577.5 | 578.6
512, 1536 | 30.6 | 25.8 | 25.6
512, 6144 | 99.0 | 80.6 | 80.6
512, 10240 | 205.4 | 128.6 | 128.5
1000, 1000 | 30.2 | 24.0 | 23.8
2000, 2000 | 81.0 | 56.4 | 56.7
10240, 10240 | 1963.8 | 1323.1 | 1326.9
384, 128 | 20.7 | 20.5 | 20.0
2048, 1024 | 45.3 | 33.0 | 33.0
267, 513 | 20.7 | 20.1 | 19.8
67, 123479 | 2322.7 | 1230.8 | 1313.2
1024, 123479 | 4244.2 | 3485.9 | 3491.3
1234154, 512 | 6838.8 | 5890.0 | 5956.4
2048, 66679 | 3304.3 | 2477.3 | 2483.4
200, 256 | 20.5 | 19.7 | 19.7
1000, 256 | 20.5 | 19.9 | 19.6
6000, 256 | 32.9 | 23.7 | 23.8
6272, 256 | 34.0 | 24.8 | 24.9
200, 512 | 21.0 | 19.7 | 20.2
1000, 512 | 20.3 | 20.2 | 19.8
6000, 512 | 56.4 | 38.4 | 38.6
6272, 512 | 59.0 | 40.3 | 40.9
200, 1024 | 20.7 | 19.7 | 20.0
1000, 1024 | 30.2 | 24.1 | 24.0
6000, 1024 | 118.6 | 67.4 | 67.9
6272, 1024 | 123.8 | 73.3 | 72.6
200, 2048 | 28.9 | 26.3 | 26.2
1000, 2048 | 52.3 | 40.5 | 40.2
6000, 2048 | 238.6 | 159.3 | 159.2
6272, 2048 | 246.8 | 165.6 | 165.7
200, 3072 | 39.9 | 36.3 | 36.2
1000, 3072 | 75.0 | 56.8 | 56.5
6000, 3072 | 352.0 | 240.9 | 241.9
6272, 3072 | 365.4 | 249.1 | 250.1
3000000, 512 | 16525.9 | 14259.5 | 14365.3
Times are in microseconds (us).
[-------------------------------------- Vectorized --------------------------------------]
| fwd, torch.float32 | fwd, torch.float16 | fwd, torch.bfloat16
1 threads: -------------------------------------------------------------------------------
32, 32 | 19.4 | 19.6 | 19.6
64, 32 | 19.6 | 19.5 | 20.1
256, 128 | 19.4 | 19.6 | 20.2
512, 1024 | 19.3 | 19.9 | 20.0
1024, 2048 | 30.2 | 29.4 | 29.3
2048, 2048 | 42.4 | 35.1 | 35.0
4096, 16384 | 613.5 | 562.1 | 564.2
70000, 64 | 50.2 | 50.1 | 50.2
131072, 512 | 548.1 | 467.3 | 470.2
1000, 520 | 19.7 | 20.3 | 19.8
4005, 4005 | 206.0 | 240.6 | 240.6
10000, 1000 | 98.3 | 75.4 | 75.1
1024, 10000 | 250.9 | 200.9 | 208.1
8192, 4096 | 317.8 | 285.0 | 286.0
10000, 10000 | 857.6 | 742.6 | 743.9
3072, 10000 | 347.1 | 320.1 | 320.0
6144, 10000 | 501.9 | 433.7 | 435.7
1024, 20000 | 486.9 | 471.7 | 471.3
512, 1536 | 21.9 | 21.7 | 21.6
512, 6144 | 64.8 | 62.7 | 62.9
512, 10240 | 139.6 | 100.2 | 100.2
1000, 1000 | 19.3 | 19.8 | 20.5
2000, 2000 | 36.5 | 34.3 | 34.3
10240, 10240 | 849.1 | 765.9 | 767.3
384, 128 | 20.0 | 20.0 | 20.1
2048, 1024 | 22.4 | 21.5 | 21.6
267, 513 | 20.6 | 20.2 | 19.7
67, 123479 | 2354.0 | 998.4 | 996.2
1024, 123479 | 3134.1 | 3418.3 | 3427.5
1234154, 512 | 5083.0 | 4251.9 | 4278.5
2048, 66679 | 2052.4 | 2323.9 | 2325.0
200, 256 | 19.7 | 19.7 | 19.9
1000, 256 | 20.0 | 19.5 | 19.6
6000, 256 | 20.2 | 19.4 | 19.7
6272, 256 | 19.9 | 19.7 | 20.2
200, 512 | 19.7 | 19.9 | 19.5
1000, 512 | 20.1 | 19.8 | 19.6
6000, 512 | 23.1 | 20.9 | 21.3
6272, 512 | 24.0 | 22.4 | 22.9
200, 1024 | 19.4 | 19.8 | 19.7
1000, 1024 | 19.7 | 19.5 | 20.1
6000, 1024 | 51.9 | 34.1 | 34.6
6272, 1024 | 55.5 | 39.0 | 37.9
200, 2048 | 24.5 | 24.4 | 24.4
1000, 2048 | 30.3 | 29.1 | 28.9
6000, 2048 | 109.2 | 96.5 | 96.9
6272, 2048 | 111.4 | 100.1 | 100.3
200, 3072 | 33.0 | 33.5 | 33.4
1000, 3072 | 41.4 | 39.5 | 39.5
6000, 3072 | 156.3 | 146.5 | 146.9
6272, 3072 | 160.9 | 150.2 | 151.0
3000000, 512 | 12385.8 | 10336.7 | 10366.0
Times are in microseconds (us).
```
Reviewed By: q10
Differential Revision: D77698275
flaviotruzzi
force-pushed
the
export-D77698275
branch
from
2fb84a9 to
f64a51e
Compare
|
@flaviotruzzi has exported this pull request. If you are a Meta employee, you can view the originating diff in D77698275. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary:
Reland D75563906 that was backed out, with fixes.
The problem was the grid was not being big enough given the config.
Further ensured vectorization, which allows 1.4Tb/s.
Test:
Reviewed By: q10
Differential Revision: D77698275