On-the-Fly Data Layout Conversion for GEMM on AI Accelerators
Bibliographic record
Abstract
GEMM accelerators used for AI typically require special layouts of their input and output data. Pre- and post-conversion to such layouts from and to standard row-major or column-major layouts degrades performance. This is particularly the case when conversion overhead cannot be amortized over multiple GEMM executions, as in, for example, LU factorization, a key computation in HPC.We propose a novel on-the-fly data layout conversion approach for GEMM used for LU factorization on Huawei’s DaVinci AI Core. The approach alleviates the bulk of conversion through the use of DMA and Vector engines to convert data layout as data moves through the memory hierarchy, and as it is processed. The approach reduces conversion overhead, but requires more use of the DMA and Vector engines, compared to when data is already in the accelerator’s required data layout.We experimentally evaluate the approach on an Ascend 910 AI processor with 32 DaVinci cores used to accelerate GEMM. We show that the approach reduces layout conversion time by 74% and that the additional use of the DMA/Vector engines reduces compute efficiency by no more than 15%. The result is an improvement in GEMM’s end-to-end performance— over explicit pre-/post-conversion—by up to ∼2X, and on average by 1.6X. In the context of LU factorization, where GEMM is repeatedly used for shrinking matrix sizes, our approach improves GEMM performance by up to 1.6X and on average by 1.4X.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.003 | 0.001 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".