tt-metal: [Blackhole Bringup] Update tests to blackhole grid/block sizes
Some tests need to get updated to use blackhole grids, and ensure correct kernels are being used. For example:
TT_METAL_DPRINT_CORES=0,0 pytest -svv tests/ttnn/unit_tests/test_to_layout.py::test_to_layout_2D[start_with_padding=False-to_layout=Layout.ROW_MAJOR-from_layout=Layout.TILE-on_device=True-width=32-height=32]
Generates 2 different kernels for different devices.
For wormhole_b0 it generates:
ttnn/cpp/ttnn/operations/data_movement/untilize/device/kernels/compute/pack_untilize.cpp
with per_core_block_cnt = 1 and per_core_block_tile_cnt=1
while for Blackhole it generates:
ttnn/cpp/ttnn/operations/data_movement/untilize/device/kernels/compute/untilize.cpp
with per_core_block_cnt = 1 and per_core_block_tile_cnt=32
The per_core_block_tile_cnt=32 is incorrect, since this test is only a single 32x32 tile, but it generates incorrectly for Blackhole device. Seems like potential issue is from here:
std::string compute_kernel("ttnn/cpp/ttnn/operations/data_movement/untilize/device/kernels/compute/pack_untilize.cpp");
if (ntiles_per_block > MAX_PACK_UNTILIZE_WIDTH || !use_pack_untilize) {
log_debug(tt::LogOp, "Using slow untilize.");
compute_kernel = std::string("ttnn/cpp/ttnn/operations/data_movement/untilize/device/kernels/compute/untilize.cpp");
} else {
log_debug(tt::LogOp, "Using fast pack untilize.");
}
All potential tests that use the above kernels should be updated for Blackhole.
About this issue
- Original URL
- State: closed
- Created 2 years ago
- Comments: 19 (19 by maintainers)
Fix is on https://github.com/tenstorrent/tt-metal/pull/11983