commit: 8b85022c8f432f77ffd9d25eda7cc6884a2fdce8
[log]
author: Zhewen Yu <zhewenyu@amd.com>
Tue Apr 21 17:24:29 2026 +0200
committer: GitHub <noreply@github.com>
Tue Apr 21 16:24:29 2026 +0100
tree: 603fbf3d4e2b1eb4fe1526a473e21d268d7bbe8b
parent: f629f2135ac709876644aeb5716d04ce15d49e44 [diff]

[Codegen] Enable DMA by default for F16/BF16 Gemm on gfx950 (#24117)

This PR flips the default DMA from false to true. Currently only BF16
GEMM has been properly benchmarked, but going to expand other supports
as a follow-up.

Overall GEMM benchmark: **+5.9%** geomean speedup if TN (LHS transposed,
RHS not transposed) cases are excluded. **+4.3%** if included.

#### Top 20 improved

| # | Op | Shape (MxNxK) | Transpose | NoDMA (us) | DMA (us) | Speedup |
|---|-----|--------------|--------|------------|----------|---------|
| 1 | mm | 7680x304x512 | NT | 14.4 | 9.7 | +32.7% |
| 2 | mm | 1280x3840x3840 | NN | 84.0 | 59.5 | +29.2% |
| 3 | mm | 24576x304x512 | NT | 33.4 | 23.8 | +28.6% |
| 4 | addmm | 256x3840x3840 | NT | 39.2 | 28.0 | +28.6% |
| 5 | mm | 1280x2304x576 | NT | 11.6 | 8.4 | +27.4% |
| 6 | mm | 1280x3840x7680 | NN | 164.4 | 119.4 | +27.3% |
| 7 | mm | 960x256x2048 | NT | 20.4 | 15.0 | +26.5% |
| 8 | mm | 4096x2048x8192 | NN | 215.5 | 159.9 | +25.8% |
| 9 | mm | 4096x2048x6144 | NN | 163.1 | 121.7 | +25.4% |
| 10 | mm | 4096x2048x4096 | NN | 112.7 | 84.4 | +25.1% |
| 11 | addmm | 5x384x384 | NT | 4.7 | 3.5 | +24.9% |
| 12 | mm | 3072x256x2048 | NT | 21.2 | 16.1 | +24.1% |
| 13 | addmm | 18928x128x512 | NT | 13.7 | 10.4 | +23.9% |
| 14 | mm | 1285x8192x2048 | NN | 76.9 | 59.4 | +22.9% |
| 15 | mm | 4112x2048x8192 | NN | 275.1 | 213.3 | +22.5% |
| 16 | mm | 256x3840x3840 | NN | 41.7 | 32.4 | +22.2% |
| 17 | mm | 1280x1152x576 | NT | 8.1 | 6.3 | +22.1% |
| 18 | mm | 4096x2048x2048 | NN | 57.0 | 44.5 | +22.0% |
| 19 | addmm | 1280x3840x3840 | NT | 83.9 | 65.5 | +22.0% |
| 20 | mm | 1280x3840x3840 | NT | 83.7 | 66.0 | +21.1% |

<details>
<summary>Reproduce commands (top 20 improved)</summary>

```
# 1. mm 7680x304x512 NT (+32.7%)
aten::mm '[[7680, 512], [512, 304]]' '["c10::BFloat16", "c10::BFloat16"]' '[[512, 1], [1, 512]]' '["", ""]'
# 2. mm 1280x3840x3840 NN (+29.2%)
aten::mm '[[1280, 3840], [3840, 3840]]' '["c10::BFloat16", "c10::BFloat16"]' '[[3840, 1], [3840, 1]]' '["", ""]'
# 3. mm 24576x304x512 NT (+28.6%)
aten::mm '[[24576, 512], [512, 304]]' '["c10::BFloat16", "c10::BFloat16"]' '[[512, 1], [1, 512]]' '["", ""]'
# 4. addmm 256x3840x3840 NT (+28.6%)
aten::addmm '[[3840], [256, 3840], [3840, 3840], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [3840, 1], [1, 3840], [], []]' '["", "", "", "1", "1"]'
# 5. mm 1280x2304x576 NT (+27.4%)
aten::mm '[[1280, 576], [576, 2304]]' '["c10::BFloat16", "c10::BFloat16"]' '[[576, 1], [1, 576]]' '["", ""]'
# 6. mm 1280x3840x7680 NN (+27.3%)
aten::mm '[[1280, 7680], [7680, 3840]]' '["c10::BFloat16", "c10::BFloat16"]' '[[7680, 1], [3840, 1]]' '["", ""]'
# 7. mm 960x256x2048 NT (+26.5%)
aten::mm '[[960, 2048], [2048, 256]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2048, 1], [1, 2048]]' '["", ""]'
# 8. mm 4096x2048x8192 NN (+25.8%)
aten::mm '[[4096, 8192], [8192, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[8192, 1], [2048, 1]]' '["", ""]'
# 9. mm 4096x2048x6144 NN (+25.4%)
aten::mm '[[4096, 6144], [6144, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[6144, 1], [2048, 1]]' '["", ""]'
# 10. mm 4096x2048x4096 NN (+25.1%)
aten::mm '[[4096, 4096], [4096, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[4096, 1], [2048, 1]]' '["", ""]'
# 11. addmm 5x384x384 NT (+24.9%)
aten::addmm '[[384], [5, 384], [384, 384], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [384, 1], [1, 384], [], []]' '["", "", "", "1", "1"]'
# 12. mm 3072x256x2048 NT (+24.1%)
aten::mm '[[3072, 2048], [2048, 256]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2048, 1], [1, 2048]]' '["", ""]'
# 13. addmm 18928x128x512 NT (+23.9%)
aten::addmm '[[128], [18928, 512], [512, 128], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [512, 1], [1, 512], [], []]' '["", "", "", "1", "1"]'
# 14. mm 1285x8192x2048 NN (+22.9%)
aten::mm '[[1285, 2048], [2048, 8192]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2048, 1], [8192, 1]]' '["", ""]'
# 15. mm 4112x2048x8192 NN (+22.5%)
aten::mm '[[4112, 8192], [8192, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[8192, 1], [2048, 1]]' '["", ""]'
# 16. mm 256x3840x3840 NN (+22.2%)
aten::mm '[[256, 3840], [3840, 3840]]' '["c10::BFloat16", "c10::BFloat16"]' '[[3840, 1], [3840, 1]]' '["", ""]'
# 17. mm 1280x1152x576 NT (+22.1%)
aten::mm '[[1280, 576], [576, 1152]]' '["c10::BFloat16", "c10::BFloat16"]' '[[576, 1], [1, 576]]' '["", ""]'
# 18. mm 4096x2048x2048 NN (+22.0%)
aten::mm '[[4096, 2048], [2048, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2048, 1], [2048, 1]]' '["", ""]'
# 19. addmm 1280x3840x3840 NT (+22.0%)
aten::addmm '[[3840], [1280, 3840], [3840, 3840], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [3840, 1], [1, 3840], [], []]' '["", "", "", "1", "1"]'
# 20. mm 1280x3840x3840 NT (+21.1%)
aten::mm '[[1280, 3840], [3840, 3840]]' '["c10::BFloat16", "c10::BFloat16"]' '[[3840, 1], [1, 3840]]' '["", ""]'
```

</details>

#### Regressions

The regression cases are strongly correlated with TN (LHS transposed,
RHS not transposed) layouts — 33 out of 83 total regressions are TN, and
they account for all of the severe cases (>25%). This PR disables DMA
for TN cases, and they will be addressed in a follow-up.

<details>
<summary>Top 20 regressed — all layouts</summary>

| # | Op | Shape (MxNxK) | Transpose | NoDMA (us) | DMA (us) |
Regression |
|---|-----|--------------|--------|------------|----------|------------|
| 1 | mm | 8192x2048x150000 | TN | 6039.3 | 14120.4 | -133.8% |
| 2 | mm | 16384x4096x150000 | TN | 24019.5 | 52548.8 | -118.8% |
| 3 | mm | 4096x16384x150000 | TN | 24946.4 | 38507.4 | -54.4% |
| 4 | mm | 2048x8192x150000 | TN | 6307.2 | 9572.7 | -51.8% |
| 5 | mm | 1024x128x150000 | TN | 190.4 | 262.3 | -37.8% |
| 6 | mm | 2048x2048x1285 | TN | 27.1 | 36.2 | -33.3% |
| 7 | mm | 128x128x2119936 | TN | 287.2 | 365.2 | -27.2% |
| 8 | mm | 576x3840x7680 | TN | 111.6 | 140.7 | -26.1% |
| 9 | mm | 1134x2048x150000 | TN | 3435.5 | 4313.4 | -25.6% |
| 10 | mm | 20x3840x21760 | TN | 42.9 | 53.7 | -25.1% |
| 11 | mm | 10x576x2304 | NN | 9.0 | 11.1 | -23.7% |
| 12 | bmm | 32x96x21x96 | NT | 4.2 | 5.1 | -23.5% |
| 13 | addmm | 5x4x384 | NT | 6.8 | 8.3 | -21.8% |
| 14 | bmm | 32x21x96x96 | NT | 4.2 | 5.1 | -21.5% |
| 15 | addmm | 150000x1134x2048 | NT | 1098.5 | 1303.2 | -18.6% |
| 16 | mm | 576x2048x24576 | TN | 167.6 | 195.8 | -16.9% |
| 17 | mm | 2048x512x24576 | TN | 142.9 | 167.0 | -16.8% |
| 18 | mm | 1024x1024x16 | TN | 4.4 | 5.1 | -16.1% |
| 19 | mm | 512x2048x24576 | TN | 136.1 | 157.9 | -16.0% |
| 20 | mm | 5x384x4 | NN | 3.5 | 4.1 | -15.2% |

<details>
<summary>Reproduce commands</summary>

```
# 1. mm 8192x2048x150000 TN (-133.8%)
aten::mm '[[8192, 150000], [150000, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 8192], [2048, 1]]' '["", ""]'
# 2. mm 16384x4096x150000 TN (-118.8%)
aten::mm '[[16384, 150000], [150000, 4096]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 16384], [4096, 1]]' '["", ""]'
# 3. mm 4096x16384x150000 TN (-54.4%)
aten::mm '[[4096, 150000], [150000, 16384]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 4096], [16384, 1]]' '["", ""]'
# 4. mm 2048x8192x150000 TN (-51.8%)
aten::mm '[[2048, 150000], [150000, 8192]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 2048], [8192, 1]]' '["", ""]'
# 5. mm 1024x128x150000 TN (-37.8%)
aten::mm '[[1024, 150000], [150000, 128]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 1024], [128, 1]]' '["", ""]'
# 6. mm 2048x2048x1285 TN (-33.3%)
aten::mm '[[2048, 1285], [1285, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 2048], [2048, 1]]' '["", ""]'
# 7. mm 128x128x2119936 TN (-27.2%)
aten::mm '[[128, 2119936], [2119936, 128]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 128], [128, 1]]' '["", ""]'
# 8. mm 576x3840x7680 TN (-26.1%)
aten::mm '[[576, 7680], [7680, 3840]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 576], [3840, 1]]' '["", ""]'
# 9. mm 1134x2048x150000 TN (-25.6%)
aten::mm '[[1134, 150000], [150000, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 1134], [2048, 1]]' '["", ""]'
# 10. mm 20x3840x21760 TN (-25.1%)
aten::mm '[[20, 21760], [21760, 3840]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 20], [3840, 1]]' '["", ""]'
# 11. mm 10x576x2304 NN (-23.7%)
aten::mm '[[10, 2304], [2304, 576]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2304, 1], [576, 1]]' '["", ""]'
# 12. bmm 32x96x21x96 (-23.5%)
aten::bmm '[[32, 96, 96], [32, 96, 21]]' '["c10::BFloat16", "c10::BFloat16"]' '[[9216, 96, 1], [2016, 1, 96]]' '["", ""]'
# 13. addmm 5x4x384 NT (-21.8%)
aten::addmm '[[4], [5, 384], [384, 4], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [384, 1], [1, 384], [], []]' '["", "", "", "1", "1"]'
# 14. bmm 32x21x96x96 (-21.5%)
aten::bmm '[[32, 21, 96], [32, 96, 96]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2016, 96, 1], [9216, 1, 96]]' '["", ""]'
# 15. addmm 150000x1134x2048 NT (-18.6%)
aten::addmm '[[1134], [150000, 2048], [2048, 1134], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [2048, 1], [1, 2048], [], []]' '["", "", "", "1", "1"]'
# 16. mm 576x2048x24576 TN (-16.9%)
aten::mm '[[576, 24576], [24576, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 576], [2048, 1]]' '["", ""]'
# 17. mm 2048x512x24576 TN (-16.8%)
aten::mm '[[2048, 24576], [24576, 512]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 2048], [512, 1]]' '["", ""]'
# 18. mm 1024x1024x16 TN (-16.1%)
aten::mm '[[1024, 16], [16, 1024]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 1024], [1024, 1]]' '["", ""]'
# 19. mm 512x2048x24576 TN (-16.0%)
aten::mm '[[512, 24576], [24576, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[1, 512], [2048, 1]]' '["", ""]'
# 20. mm 5x384x4 NN (-15.2%)
aten::mm '[[5, 4], [4, 384]]' '["c10::BFloat16", "c10::BFloat16"]' '[[4, 1], [384, 1]]' '["", ""]'
```

</details>

</details>

<details>
<summary>Top 20 regressed — excluding TN layouts</summary>

| # | Op | Shape (MxNxK) | Transpose | NoDMA (us) | DMA (us) |
Regression |
|---|-----|--------------|--------|------------|----------|------------|
| 1 | mm | 10x576x2304 | NN | 9.0 | 11.1 | -23.7% |
| 2 | bmm | 32x96x21x96 | NT | 4.2 | 5.1 | -23.5% |
| 3 | addmm | 5x4x384 | NT | 6.8 | 8.3 | -21.8% |
| 4 | bmm | 32x21x96x96 | NT | 4.2 | 5.1 | -21.5% |
| 5 | addmm | 150000x1134x2048 | NT | 1098.5 | 1303.2 | -18.6% |
| 6 | mm | 5x384x4 | NN | 3.5 | 4.1 | -15.2% |
| 7 | mm | 1285x2048x8192 | NN | 154.9 | 178.0 | -14.9% |
| 8 | mm | 1285x2048x3840 | NT | 65.0 | 74.0 | -13.9% |
| 9 | mm | 16800000x134x128 | NN | 3879.0 | 4408.3 | -13.6% |
| 10 | mm | 1285x2048x2048 | NN | 43.2 | 48.9 | -13.1% |
| 11 | mm | 1285x2048x3840 | NN | 76.2 | 86.1 | -12.9% |
| 12 | mm | 24576x512x2048 | NN | 102.8 | 115.2 | -12.1% |
| 13 | bmm | 16x384x384x192 | NT | 6.4 | 7.1 | -10.9% |
| 14 | bmm | 32x96x96x21 | NN | 4.0 | 4.4 | -10.8% |
| 15 | bmm | 32x96x96x96 | NT | 4.5 | 5.0 | -10.4% |
| 16 | bmm | 32x96x96x96 | NN | 4.5 | 4.9 | -9.4% |
| 17 | mm | 528x2304x576 | NN | 9.2 | 10.1 | -8.9% |
| 18 | bmm | 32x96x96x21 | NN | 6.5 | 7.0 | -8.6% |
| 19 | addmm | 1285x2048x2048 | NT | 36.9 | 39.9 | -8.0% |
| 20 | mm | 24576x2048x512 | NN | 106.6 | 114.7 | -7.7% |

<details>
<summary>Reproduce commands</summary>

```
# 1. mm 10x576x2304 NN (-23.7%)
aten::mm '[[10, 2304], [2304, 576]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2304, 1], [576, 1]]' '["", ""]'
# 2. bmm 32x96x21x96 (-23.5%)
aten::bmm '[[32, 96, 96], [32, 96, 21]]' '["c10::BFloat16", "c10::BFloat16"]' '[[9216, 96, 1], [2016, 1, 96]]' '["", ""]'
# 3. addmm 5x4x384 NT (-21.8%)
aten::addmm '[[4], [5, 384], [384, 4], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [384, 1], [1, 384], [], []]' '["", "", "", "1", "1"]'
# 4. bmm 32x21x96x96 (-21.5%)
aten::bmm '[[32, 21, 96], [32, 96, 96]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2016, 96, 1], [9216, 1, 96]]' '["", ""]'
# 5. addmm 150000x1134x2048 NT (-18.6%)
aten::addmm '[[1134], [150000, 2048], [2048, 1134], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [2048, 1], [1, 2048], [], []]' '["", "", "", "1", "1"]'
# 6. mm 5x384x4 NN (-15.2%)
aten::mm '[[5, 4], [4, 384]]' '["c10::BFloat16", "c10::BFloat16"]' '[[4, 1], [384, 1]]' '["", ""]'
# 7. mm 1285x2048x8192 NN (-14.9%)
aten::mm '[[1285, 8192], [8192, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[8192, 1], [2048, 1]]' '["", ""]'
# 8. mm 1285x2048x3840 NT (-13.9%)
aten::mm '[[1285, 3840], [3840, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[3840, 1], [1, 3840]]' '["", ""]'
# 9. mm 16800000x134x128 NN (-13.6%)
aten::mm '[[16800000, 128], [128, 134]]' '["c10::BFloat16", "c10::BFloat16"]' '[[128, 1], [134, 1]]' '["", ""]'
# 10. mm 1285x2048x2048 NN (-13.1%)
aten::mm '[[1285, 2048], [2048, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2048, 1], [2048, 1]]' '["", ""]'
# 11. mm 1285x2048x3840 NN (-12.9%)
aten::mm '[[1285, 3840], [3840, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[3840, 1], [2048, 1]]' '["", ""]'
# 12. mm 24576x512x2048 NN (-12.1%)
aten::mm '[[24576, 2048], [2048, 512]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2048, 1], [512, 1]]' '["", ""]'
# 13. bmm 16x384x384x192 NT (-10.9%)
aten::bmm '[[16, 384, 192], [16, 192, 384]]' '["c10::BFloat16", "c10::BFloat16"]' '[[73728, 192, 1], [73728, 1, 192]]' '["", ""]'
# 14. bmm 32x96x96x21 NN (-10.8%)
aten::bmm '[[32, 96, 21], [32, 21, 96]]' '["c10::BFloat16", "c10::BFloat16"]' '[[2016, 21, 1], [2016, 96, 1]]' '["", ""]'
# 15. bmm 32x96x96x96 NT (-10.4%)
aten::bmm '[[32, 96, 96], [32, 96, 96]]' '["c10::BFloat16", "c10::BFloat16"]' '[[9216, 96, 1], [9216, 1, 96]]' '["", ""]'
# 16. bmm 32x96x96x96 NN (-9.4%)
aten::bmm '[[32, 96, 96], [32, 96, 96]]' '["c10::BFloat16", "c10::BFloat16"]' '[[9216, 96, 1], [9216, 96, 1]]' '["", ""]'
# 17. mm 528x2304x576 NN (-8.9%)
aten::mm '[[528, 576], [576, 2304]]' '["c10::BFloat16", "c10::BFloat16"]' '[[576, 1], [2304, 1]]' '["", ""]'
# 18. bmm 32x96x96x21 NN (-8.6%)
aten::bmm '[[32, 96, 21], [32, 21, 96]]' '["c10::BFloat16", "float"]' '[[2016, 21, 1], [2016, 96, 1]]' '["", ""]'
# 19. addmm 1285x2048x2048 NT (-8.0%)
aten::addmm '[[2048], [1285, 2048], [2048, 2048], [], []]' '["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar"]' '[[1], [2048, 1], [1, 2048], [], []]' '["", "", "", "1", "1"]'
# 20. mm 24576x2048x512 NN (-7.7%)
aten::mm '[[24576, 512], [512, 2048]]' '["c10::BFloat16", "c10::BFloat16"]' '[[512, 1], [2048, 1]]' '["", ""]'
```

</details>

</details>

---------

Signed-off-by: Yu-Zhewen <zhewenyu@amd.com>

6 files changed

tree: 603fbf3d4e2b1eb4fe1526a473e21d268d7bbe8b

README.md

IREE: Intermediate Representation Execution Environment

IREE (Intermediate Representation Execution Eenvironment, pronounced as “eerie”) is an MLIR-based end-to-end compiler and runtime that lowers Machine Learning (ML) models to a unified IR that scales up to meet the needs of the datacenter and down to satisfy the constraints and special considerations of mobile and edge deployments.

See our website for project details, user guides, and instructions on building from source.

Project news

Project status

Release status

Releases notes are published on GitHub releases.

Package	Release status
GitHub release (stable)
GitHub release (nightly)
`iree-base-compiler`
`iree-base-runtime`

For more details on the release process, see https://iree.dev/developers/general/release-management/.

Build status

Nightly build status

Operating system	Build status
Linux
macOS
macOS

For the full list of workflows see https://iree.dev/developers/general/github-actions/.

Communication channels

GitHub issues: Feature requests, bugs, and other work tracking
IREE Discord server: Daily development discussions with the core team and collaborators
(New) iree-announce email list: Announcements
(New) iree-technical-discussion email list: General and low-priority discussion
(Legacy) iree-discuss email list: Announcements, general and low-priority discussion

Related project channels

MLIR topic within LLVM Discourse: IREE is enabled by and heavily relies on MLIR. IREE sometimes is referred to in certain MLIR discussions. Useful if you are also interested in MLIR evolution.

Architecture overview

IREE Architecture

See our website for more information.

Presentations and talks

Community meeting recordings: IREE YouTube channel

Date	Title	Recording	Slides
2025-06-10	Data-Tiling in IREE: Achieving High Performance Through Compiler Design (AsiaLLVM)	recording	slides
2025-05-17	Introduction to GPU architecture and IREE's GPU CodeGen Pipeline	recording	slides
2025-02-12	The Long Tail of AI: SPIR-V in IREE and MLIR (Vulkanised)	recording	slides
2024-10-01	Unveiling the Inner Workings of IREE: An MLIR-Based Compiler for Diverse Hardware	recording
2021-06-09	IREE Runtime Design Tech Talk	recording	slides
2020-08-20	IREE CodeGen (MLIR Open Design Meeting)	recording	slides
2020-03-18	Interactive HAL IR Walkthrough	recording
2020-01-31	End-to-end MLIR Workflow in IREE (MLIR Open Design Meeting)	recording	slides

License

IREE is licensed under the terms of the Apache 2.0 License with LLVM Exceptions. See LICENSE for more information.