Investigate cases in which subgroup size 16 is noticeable slower

[victor-eds]

After running huggingface benchmarks with subgroup size 16 (https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/9499093865), we saw some cases in which subgroup size 16 reported worse performance:

• huggingface amp_bf16 inference PLBartForCausalLM: 3.58x vs 2.36x speedups
• huggingface bf16 inferece BartForCausalLM: 2.52x vs 1.19x speedups
• huggingface f32 training T5Small: 0.65x vs 0.50x speedups

Investigate and create followup issues if needed or write report in this issue.

Env:

TIMESTAMP=20240613123916
JOB_NAME=huggingface-inference-amp_bf16
GITHUB_RUN_ID=9499093865
GITHUB_RUN_NUMBER=102
GITHUB_RUN_ATTEMPT=1
PYTHON_VERSION=3.10
PYTORCH_REPO=Stonepia/pytorch
PYTORCH_COMMIT_ID=22ce6c6508d1d13b263d4c8b1fd6b98505983e92
IPEX_REPO=
IPEX_COMMIT_ID=
LLVM_REPO=llvm/llvm-project
LLVM_COMMIT_ID=765206e050453018e861637a08a4520f29238074
BENCHMARK_REPO=weishi-deng/benchmark.git
BENCHMARK_COMMIT_ID=e90564f719a8df7daf0cff4a245404c435b7693a
TRITON_REPO=intel/intel-xpu-backend-for-triton
TRITON_COMMIT_ID=b25bf436956b8fce31bdd8dba68a9aa3959917e6
TORCHVISION_COMMIT_ID=
TORCHTEXT_COMMIT_ID=
TORCHAUDIO_COMMIT_ID=
TRANSFORMERS_VERSION=4.27.4
TIMM_COMMIT_ID=b9d43c7dcac1fe05e851dd7be7187b108af593d2
LIBIGC1_VERSION=1.0.16510.19-881
LEVEL_ZERO_VERSION=1.3.29138.29-881
GPU_DEVICE=Intel(R) Data Center GPU Max 1550
AGAMA_VERSION=881

[sommerlukas]

So far, we haven't been able to reproduce this behavior on the current llvm-target branch. Will investigate further.

[sommerlukas]

Work on this ticket had been blocked by #1647 for most of last week. We were able to obtain data from a CI run end of last week and will present a summary offline in the meeting.

[sommerlukas]

As discussed in chat offline, varying performance between different CI runs (based on the same commit) are still a problem for this investigation.

We'll try to still identify a consistent outlier in the performance comparison with the different subgroups sizes and then investigate what causes that performance difference.

[sommerlukas]

After comparing CI runs 4 and 5, three outlier models could be identified where sub-group size 16 consistently across both runs provided worse performance than sub-group size 32:

• AllenaiLongformerBase training with float16 and amp_fp16
• XLNetLMHeadModel training with float16 and amp_fp16
• BlenderbotSmallForCausalLM inference with amp_fp16

For XLNetLMHeadModel, neither sub-group size provides a speedup over Pytorch execution, so the model was excluded from further investigation.

AllenaiLongformerBase

Comparing device timing from unitrace for both SG-sizes shows two Triton kernels that are among the GPU kernels that take up the most time.

For SG-size 32:

                                   Kernel,        Calls,            Time (ns),     Time (%),         Average (ns),             Min (ns),             Max (ns)
"triton_poi_fused_index_add_new_zeros_13",          179,            480605600,     8.665852,              2684947,                28800,              2943360
"triton_poi_fused_index_add_new_zeros_25",          178,            443019040,     7.988125,              2488871,                 6080,              2830400

For SG-size 16:

                                   Kernel,        Calls,            Time (ns),     Time (%),         Average (ns),             Min (ns),             Max (ns)
"triton_poi_fused_index_add_new_zeros_13",          178,            631492960,    10.590560,              3547713,                 2720,              4099520
"triton_poi_fused_index_add_new_zeros_25",          178,            629485440,    10.556891,              3536435,                 6080,              4015200

Whereas these two kernels each take up ~8% of the overall execution time of the model with SG-size 32, they take up 10.5% with SG-size 16.

The average execution time of the kernel also increases from 2.6ms to 3.5ms, a 1.3x slowdown.

The kernels are attached to this comment. They are both rather simple, but both use tl.atomic_add. @chengjunlu had confirmed in a chat offline that performance for atomic operations is a known issue, so that the performance difference potentially stems from this operation.

BlenderbotSmallForCausalLM

Comparing device timing from unitrace for both SG-sizes shows one Triton kernel that is among the GPU kernels that take up the most time.

For SG-size 32:

                                   Kernel,        Calls,            Time (ns),     Time (%),         Average (ns),             Min (ns),             Max (ns)
"triton_red_fused__log_softmax__to_copy_view_8",           15,             42188320,     6.827563,              2812554,              2708160,              2900960

For SG-size 16:

                                   Kernel,        Calls,            Time (ns),     Time (%),         Average (ns),             Min (ns),             Max (ns)
"triton_red_fused__log_softmax__to_copy_view_8",           15,             60684640,     9.484119,              4045642,              3979680,              4196800

Whereas the kernel takes up less than 7% of the overall execution time of the model with SG-size 32, it takes up 9.5% with SG-size 16.

The average execution time of the kernel also increases from 2.8ms to 4ms, a 1.43x slowdown.

The kernel is attached to this comment and is also rather simple. It however uses a reduction. From previous investigations by @victor-eds, it is known that the current pattern used for reductions in the XPU backend is less efficient for SG-size 16, as it generates significantly more assembly instructions, so the performance difference most likely stems from this known issue.

add_new_zeros_25.txt
add_new_zeros_13.txt
softmax_to_copy_view_8.txt

[sommerlukas]

I filed #1867 and #1868 as follow-up to investigate the root cause of the two outliers. That investigation currently has lower priority.