Flash attention 140 TFlops builder with barriers removal and experimental TILE_S1=512 version

[MirkoDeVita98]

The following performance is measured with:

bash compile.sh --remove-vec-barriers 1264,1267,1272,1275,1279,1282,1311,1313,1316,1320,1322,1325,1328,1330,1333,1362,1364,1367,1371,1373,1376,1379,1381,1384,1390
python run.py

[MirkoDeVita98]

2026/05/13

Commit 5176daa

Examples

$ python examples/validate_all_examples.py 
FAILED aot/flash_attention/experimental [0.00s]
PASSED aot/activations/geglu_dynamic_multicore [37.29s]
PASSED aot/activations/relu_dynamic_multicore [29.13s]
PASSED aot/batch_matmul/matmul_dynbatch_multicore [22.61s]
PASSED aot/batch_matmul/matmul_dynbatch_multicore_2buf [19.84s]
PASSED aot/batch_matmul/matmul_dynbatch_multicore_opt [25.23s]
PASSED aot/elementwise/add_dynamic_multicore [41.64s]
PASSED aot/fast_hadamard [112.13s]
PASSED aot/fast_inverse/basic_dense [66.37s]
PASSED aot/fast_inverse/block_inversion [44.29s]
PASSED aot/flash_attention/140tflops [27.42s]
PASSED aot/flash_attention [44.51s]
PASSED aot/matmul_optimization_guide [163.03s]
PASSED aot/matmul_optimization_guide/experimental [74.86s]
PASSED aot/print_tile [24.72s]
PASSED aot/simple_static/add_static_multicore [22.02s]
PASSED aot/simple_static/matmul_static_singlecore [21.51s]
PASSED aot/sinkhorn_demo [19.06s]
PASSED aot/topk [53.56s]
PASSED aot/tpushpop/mix-kernel_mlir [83.47s]
PASSED jit/add_dynamic_multicore [19.91s]
PASSED jit/add_static_multicore [38.26s]
PASSED jit/matmul_dynamic_multicore [21.08s]
PASSED jit/scan [21.20s]
==============================================================================
23 passed, 1 failed in 1033.18s

Unit tests

$ pytest -v ./tests
======================================================= 615 passed, 21 skipped in 464.11s (0:07:44) ========================================================