DPRINT asserts hardcoded L1 address check causing dprint failure on WH_B0

[yugaoTT]

is the CB is allocated above 1024 * 1024 or below 120*1024, the dprint failed with an error BAD TILE POINTER count=0.

[tt-dma]

Could you provide an example that reproduces this bug? So we can check if the bounds need updating or if there's a bug in the DPRINT addressing

[yugaoTT]

yes,
branch yugao/moreh-u-benchmark
make build
make tests
run ./build/test/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm --m 32 --n 32 --k 64 --dtype 1 --fidel 0 --block 1 --num-tests 1 --fast-dispatch --bypass-check --one-core 1 --num-blocks 1

[tt-dma]

Are there additional defines that are needed to run this test? I'm getting:

(python_env) davidma@e04cs04:~/tt-metal$ ./build/test/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm --m 32 --n 32 --k 64 --dtype 1 --fidel 0 --block 1  --num-tests 1 --fast-dispatch --bypass-check --one-core 1 --num-blocks 1
Remaining test_args:
        --dtype
        1
        --num-blocks
        1
                   Test | ERROR    | Command line arguments found exception
                 Always | ERROR    | In the slow dispatch mode, device profiler is used to measure the performance. Build the Metal library and test code with 'ENABLE_PROFILER=1'
                   Test | ERROR    | TT_ASSERT @ tests/tt_metal/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm.cpp:199: false
backtrace:
 --- /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf3) [0x7f6b9839e083]
 --- ./build/test/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm(+0xa0ae) [0x55a4ed47f0ae]

                   Test | ERROR    | System error message: Success
                   Test | ERROR    | Test Failed

Seems it needs either profiler enabled or a certain dispatch mode?

[yugaoTT]

Which branch are you working on? David Ma ***@***.***>于2024年3月4日 周一17:10写道：

…

Are there additional defines that are needed to run this test? I'm getting: (python_env) ***@***.***:~/tt-metal$ ./build/test/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm --m 32 --n 32 --k 64 --dtype 1 --fidel 0 --block 1 --num-tests 1 --fast-dispatch --bypass-check --one-core 1 --num-blocks 1 Remaining test_args: --dtype 1 --num-blocks 1 Test | ERROR | Command line arguments found exception Always | ERROR | In the slow dispatch mode, device profiler is used to measure the performance. Build the Metal library and test code with 'ENABLE_PROFILER=1' Test | ERROR | TT_ASSERT @ tests/tt_metal/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm.cpp:199: false backtrace: --- /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf3) [0x7f6b9839e083] --- ./build/test/tt_metal/perf_microbenchmark/1_compute_mm/test_compute_mm(+0xa0ae) [0x55a4ed47f0ae] Test | ERROR | System error message: Success Test | ERROR | Test Failed Seems it needs either profiler enabled or a certain dispatch mode? — Reply to this email directly, view it on GitHub <#5980 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BCWBFVB7JMAWFBBPQP2KIR3YWTWPDAVCNFSM6AAAAABEFUFJ2GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZXGU2TKMRTGA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[tt-dma]

I checked out origin/yugao/moreh-u-benchmark

[yugaoTT]

Hmm, maybe try make build ENABLE_PROFILER=1, then make tests ENABLE_PROFILER=1. and commit should be at e7d6eaf David Ma ***@***.***>于2024年3月4日 周一17:37写道：

…

I checked out origin/yugao/moreh-u-benchmark — Reply to this email directly, view it on GitHub <#5980 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BCWBFVAEBFCQEEXDY6QEOKLYWTZT5AVCNFSM6AAAAABEFUFJ2GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZXGU4TAMBYGI> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[yugaoTT]

Oh sorry you are trying to dprint. Then it shouldn't be built with enable profiles. make build make test should work. If it doesn't, I will double check after went back home

[tt-dma]

I'm running on an n300 BM, and I'm able to run your test after checking out that commit (no extra env vars). I see a bunch of prints and then this error:

index: 2046 1 1
index: 2047 1 1
                   Test | ERROR    | validation single core failed
                  Metal | INFO     | Closing device 0
                 Always | INFO     | CSV_MICROBENCHMARK:title:test_compute_mm
                 Always | INFO     | CSV_INPUT:M:32:N:32:K:64:fast-dispatch:true

This looks separate from the DPRINT error that would be tripped by those bounds?

[yugaoTT]

Right, this is not related to dprint. David Ma ***@***.***>于2024年3月4日 周一17:53写道：

…

I'm running on an n300 BM, and I'm able to run your test after checking out that commit (no extra env vars). I see a bunch of prints and then this error: index: 2046 1 1 index: 2047 1 1 Test | ERROR | validation single core failed Metal | INFO | Closing device 0 Always | INFO | CSV_MICROBENCHMARK:title:test_compute_mm Always | INFO | CSV_INPUT:M:32:N:32:K:64:fast-dispatch:true This looks separate from the DPRINT error that would be tripped by those bounds? — Reply to this email directly, view it on GitHub <#5980 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BCWBFVDJNVSVG47T4DFGHT3YWT3NNAVCNFSM6AAAAABEFUFJ2GVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTSNZXGYYDQOBTHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[tt-dma]

Any idea on how I can reproduce the dprint issue that is filed here then?

[yugaoTT]

please enable the printing on core 0,0. export TT_METAL_DPRINT_CORES="(0,0)", the DPRINT is in file tests/tt_metal/tt_metal/perf_microbenchmark/1_compute_mm/kernels/bmm_large_block_zm_fused_bias_activation_block.cpp, at line 115