Slow convolutions on CPU with autocast

Aknw_Fen · December 14, 2024, 4:09pm

Context

After observing slower training (by logging.) using autocast, a profiling was run to check for expensive operations.

This shows CPU results, but using T4s (GPU) in Colab, bfloat16 takes very long (just like float16 does in the CPU below.). However, float16 does run faster than float32. (Apparently the reason for this is that T4s do not support bfloat16. I’m unsure if this is correct, but seems the case in Colab.)

But the issue here is mainly for CPUs.

Hardware: CPU - 11th Gen Intel(R) Core™ i7-1165G7 @ 2.80GHz (aka Intel CORE i7 11th)

Results

Running a profiler on inference (on training mode) using an autoencoder, with a snippet like:

    with (
      profile(
        activities=[
          ProfilerActivity.CPU,
        ],
        record_shapes=True,
        profile_memory=True,
      ) as prof,
      torch.autocast(get_device_name(), torch.float16),
      record_function("model_inference"),
    ):
      model(img)
    print(prof.key_averages().table(sort_by="cpu_time_total", row_limit=3))

Using 500 images on a network of 10Mb weights

It outputs, for each case, sorted by total cpu time:

float16

DEBUG:root:Device cpu 
INFO:root:--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                            Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg       CPU Mem  Self CPU Mem    # of Calls  
--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                    aten::linear         0.00%      49.990us       159.06%        2.392s     597.946ms      46.11 Mb           0 b             4  
                 model_inference         0.16%       2.472ms       100.00%        1.504s        1.504s       7.44 Mb    -250.91 Mb             1  
                     aten::addmm        79.36%        1.193s        79.49%        1.195s     597.630ms      19.34 Mb      19.34 Mb             2  
--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
Self CPU time total: 1.504s

float 32:

DEBUG:root:Device cpu
INFO:root:--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                            Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg       CPU Mem  Self CPU Mem    # of Calls  
--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                 model_inference         4.88%       8.560ms       100.00%     175.375ms     175.375ms           0 b    -495.96 Mb             1  
               aten::convolution         0.06%     107.845us        52.11%      91.393ms      15.232ms     107.55 Mb           0 b             6  
              aten::_convolution        32.91%      57.722ms        52.05%      91.285ms      15.214ms     107.55 Mb           0 b             6  
--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
Self CPU time total: 175.375ms

bfloat16

DEBUG:root:Device cpu
INFO:root:--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
                            Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg       CPU Mem  Self CPU Mem    # of Calls  
--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
          aten::conv_transpose2d         0.02%      58.300us       119.44%     453.313ms      75.552ms      37.46 Mb           0 b             6  
                 model_inference         0.90%       3.428ms       100.00%     379.540ms     379.540ms       7.44 Mb    -250.91 Mb             1  
               aten::convolution         0.03%     123.037us        71.21%     270.288ms      45.048ms      53.78 Mb           0 b             6  
--------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  
Self CPU time total: 379.540ms

Both conv and transposed convolution should be faster in half precision, but they are slower than float. Any ideas what it going wrong?

Aknw_Fen · December 14, 2024, 4:17pm

That was run within a Docker container, but in a standard system install the results are almost the same.

Aknw_Fen · December 14, 2024, 5:25pm

Just forgot that i commented in a related post and gh issue here Autocast on cpu dramatically slow - #3 by Laurent1

No solution yet apparently. I’ll test what happens on T4s…

Also, issue not be autocast but bf16, related issue below:

github.com/pytorch/pytorch

torch.autocast() hangs on CPUs

opened 03:46AM - 18 Oct 23 UTC

closed 03:27AM - 21 Feb 24 UTC

YunfanZhang42

module: cpu triaged module: amp (automated mixed precision)

### 🐛 Describe the bug Hi PyTorch Team, First and foremost, thank you for y…our invaluable contributions to the ML community! I recently encountered a performance issue when using torch.autocast() on CPUs. In the FP32 mode, everything is fine. When turning on the AMP with BF16, the code seems to be stuck and never finishes. Here is a quick repo: ```python3 import time import torch from transformers import RobertaForSequenceClassification, RobertaTokenizer SEQUENCE_LENGTH = 512 def generate_random_batch(tokenizer, batch_size, sequence_length=SEQUENCE_LENGTH): """Generate a batch of random sequences for testing.""" return tokenizer( [" ".join(["I am a test string."] * sequence_length) for _ in range(batch_size)], padding="max_length", truncation=True, max_length=SEQUENCE_LENGTH, return_tensors="pt", ) def benchmark_throughput(model, tokenizer, mixed_precision=False, batch_size=32, num_iterations=3): """Test and print the throughput of the model.""" input_data = generate_random_batch(tokenizer, batch_size).to(device) # Warm up for _ in range(3): with torch.no_grad(): # PyTorch only supports bfloat16 mixed precision on CPUs. with torch.autocast(device_type="cpu", dtype=torch.bfloat16, enabled=mixed_precision): _ = model(**input_data) start_time = time.time() for _ in range(num_iterations): with torch.no_grad(): with torch.autocast(device_type="cpu", dtype=torch.bfloat16, enabled=mixed_precision): _ = model(**input_data) end_time = time.time() elapsed_time = end_time - start_time sequences_per_second = (batch_size * num_iterations) / elapsed_time latency = 1000 / sequences_per_second * batch_size return sequences_per_second, latency if __name__ == "__main__": device = torch.device("cpu") tokenizer = RobertaTokenizer.from_pretrained("roberta-base") model = RobertaForSequenceClassification.from_pretrained("roberta-base").to(device).eval() model = torch.compile(model) throughput, latency = benchmark_throughput(model, tokenizer, mixed_precision=False) print(f"FP32 Throughput: {throughput:.2f} sequences/second, Latency: {latency:.2f} ms") throughput, latency = benchmark_throughput(model, tokenizer, mixed_precision=True) print(f"Mixed Precision Throughput: {throughput:.2f} sequences/second, Latency: {latency:.2f} ms") ``` On my Intel i7-13700K, FP 32 throughput is 4.98 sequences/second. Enabling AMP with autocast seems to hang the code, as 1 core spikes to 100% and the script would not finish after waiting for 30 minutes. Would appreciate a further look into this issue - thanks in advance! ### Versions Collecting environment information... PyTorch version: 2.1.0+cu121 Is debug build: False CUDA used to build PyTorch: 12.1 ROCM used to build PyTorch: N/A OS: Ubuntu 22.04.3 LTS (x86_64) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 Clang version: Could not collect CMake version: version 3.27.6 Libc version: glibc-2.35 Python version: 3.10.12 (main, Jun 11 2023, 05:26:28) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-6.2.0-34-generic-x86_64-with-glibc2.35 Is CUDA available: True CUDA runtime version: 11.5.119 CUDA_MODULE_LOADING set to: LAZY GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090 Nvidia driver version: 535.113.01 cuDNN version: Could not collect HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True CPU: Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit Address sizes: 46 bits physical, 48 bits virtual Byte Order: Little Endian CPU(s): 24 On-line CPU(s) list: 0-23 Vendor ID: GenuineIntel Model name: 13th Gen Intel(R) Core(TM) i7-13700K CPU family: 6 Model: 183 Thread(s) per core: 2 Core(s) per socket: 16 Socket(s): 1 Stepping: 1 CPU max MHz: 5400.0000 CPU min MHz: 800.0000 BogoMIPS: 6835.20 Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi umip pku ospke waitpkg gfni vaes vpclmulqdq tme rdpid movdiri movdir64b fsrm md_clear serialize pconfig arch_lbr ibt flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 640 KiB (16 instances) L1i cache: 768 KiB (16 instances) L2 cache: 24 MiB (10 instances) L3 cache: 30 MiB (1 instance) NUMA node(s): 1 NUMA node0 CPU(s): 0-23 Vulnerability Gather data sampling: Not affected Vulnerability Itlb multihit: Not affected Vulnerability L1tf: Not affected Vulnerability Mds: Not affected Vulnerability Meltdown: Not affected Vulnerability Mmio stale data: Not affected Vulnerability Retbleed: Not affected Vulnerability Spec rstack overflow: Not affected Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Versions of relevant libraries: [pip3] numpy==1.26.1 [pip3] pytorch-lightning==2.1.0 [pip3] torch==2.1.0 [pip3] torchaudio==2.1.0 [pip3] torchmetrics==1.2.0 [pip3] torchvision==0.16.0 [pip3] triton==2.1.0 [conda] Could not collect cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @mcarilli @ptrblck @leslie-fang-intel