#10548: Support tile layout for width/height-sharded concat

[jerrysky3]

Ticket

#10548

Problem description

Change kernel to support concat on both width and height-sharded tensors in tile layout. The kernel supports the following cases:

• Height-sharded width concat in row-major/tile layout
• Width-sharded height concat in row-major/tile layout

For now it's only used to concat > 2 tensors. Two tensor concatenation is currently handled by a special unrolled kernel due to performance issue with runtime args. This new kernel can be unrolled and replace it if needed in the future

What's changed

• Rename and change s2s_rm_concat_multi_core to s2s_concat_multi_core to support tile layout besides the row-major layout
• Rename reader_s2s_rm_tensor_concat.cpp to reader_s2s_tensor_concat.cpp

Checklist

• Post commit CI passes https://github.com/tenstorrent/tt-metal/actions/runs/11396048642
• Blackhole Post commit (if applicable)
• Model regression CI testing passes (if applicable) https://github.com/tenstorrent/tt-metal/actions/runs/11396057538
• Device performance regression CI testing passes (if applicable) https://github.com/tenstorrent/tt-metal/actions/runs/11396055725, the fails are due to other changes
• New/Existing tests provide coverage for changes

[jerrysky3]

Device performance regression (https://github.com/tenstorrent/tt-metal/actions/runs/11396055725) is failing with:

AssertionError: Some model(s) AVG DEVICE KERNEL SAMPLES/S are faster than expected, see above for details. {'AVG DEVICE KERNEL SAMPLES/S': [('ttnn_distilbert8_distilbert-base-uncased-distilled-squad', 90.4624, 20.394), ('ttnn_functional_ttnn_vgg11_1_', 270.5544, 108.871), ('ttnn_functional_ttnn_vgg16_1_', 195.0345, 95.378)]}

However it is currently also happening on the main branch: https://github.com/tenstorrent/tt-metal/actions/runs/11399044245/job/31717674624#step:9:2739

[sjameelTT]

Looks good to me, nice work.

[sjameelTT]

Device performance regression (https://github.com/tenstorrent/tt-metal/actions/runs/11396055725) is failing with:

AssertionError: Some model(s) AVG DEVICE KERNEL SAMPLES/S are faster than expected, see above for details. {'AVG DEVICE KERNEL SAMPLES/S': [('ttnn_distilbert8_distilbert-base-uncased-distilled-squad', 90.4624, 20.394), ('ttnn_functional_ttnn_vgg11_1_', 270.5544, 108.871), ('ttnn_functional_ttnn_vgg16_1_', 195.0345, 95.378)]}

However it is currently also happening on the main branch: https://github.com/tenstorrent/tt-metal/actions/runs/11399044245/job/31717674624#step:9:2739

Yeah this was actually me, perf numbers have improved on a bunch of models but I didn't update them with my change since we fail on very big improvements too (to prevent the numbers from getting stale).

[jerrysky3]

Hi @ayerofieiev-tt , this is another kernel ramp-up task that needs a code owner review and merge. Thanks!