Performance testing of convolutions on Radeon VII: all solvers vs deprecated solvers disabled.

[atamazov]

An investigation of an imaginary case where several old OpenCL-based solvers are turned off.

Prerequisites

• gfx906/60
• ROCm 5.4.1
• MIOpen release build
  • DEV=Off to have binary cache
  • COMgr and HIPRTC are disabled to enable parallel builds (for performance of tuning)
• The configurations under test are fully tuned locally:
  • MIOPEN_FIND_ENFORCE=3 (this fills the omissions in perf-db, if any)
  • The binary kernel cache is filled with necessary kernels
  • The User Find-Db is also updated.

Results

• All solvers enabled vs deprecated solvers disabled
  • FP32 GPU time 0.85 (average perf.gain), 0.51 (sum of times)
  • FP16 GPU time 0.81 (average perf.gain), 0.31 (sum of times)
• The same but iGEMM ASM solvers constantly disabled (no MISA-generated kernels)
  • FP32 GPU time 0.84 (average perf.gain), 0.48 (sum of times)
  • FP16 GPU time 0.81 (average perf.gain), 0.32 (sum of times)

The most affected convolution is WrW

• fp32 0.76 (avg) 0.32 (sum)
• fp16 0.58 (avg) 0.18 (sum)

The least affected convolution is Forward -- its performance drop is less than 5%.

Bottom line

Most likely, inference won't be affected, but I am expecting ~2X perf drop for FP32 training and possibly even worse for FP16 training.

---

CC @junliume @JehandadKhan

• https://github.com/ROCmSoftwarePlatform/MIOpen/labels/performance