1. tools

  • nsight in WIN（vs） or Linux (eclipse）
  • nvprof in linux cmd line 
    
//in gtx1060 
nvprof --metrics ipc,issued_ipc,achieved_occupancy,global_hit_rate,local_hit_rate,l2_tex_read_hit_rate,gld_transactions,gst_transactions,local_load_transactions,local_store_transactions,l2_tex_read_transactions,l2_tex_write_transactions,l2_read_transactions,l2_write_transactions,dram_read_transactions,dram_write_transactions,sysmem_read_transactions,sysmem_write_transactions ./wave

2. 度量标准 metrics

2.1 Performance

  • ipc
    • Instructions executed per cycle
  • issued_ipc
    • Instructions issued per cycle
  • achieved_occupancy
    • Ratio of the average active warps per active cycle to the maximum number of warps supported on a multiprocessor

说明：本文研究点在 Data Cache，那么一下的提到的L1 Cache 都为 Data Cache

2.2 Cache Hit Rate

L1 Cache

Fermi/Kepler (Capability 2.x/3.x)

  • l1_cache_global_hit_rate
    • Hit rate in L1 cache for global loads
  • l1_cache_local_hit_rate
    • Hit rate in L1 cache for local loads and stores
  • nc_cache_global_hit_rate
    • only for Kepler
    • Hit rate in non coherent cache for global loads

Maxwell/Pascal(Capability 5.x/6.x)

  • global_hit_rate
    • Hit rate for global loads
  • local_hit_rate
    • Hit rate for local loads and stores

L2 Cache

Fermi/Kepler (Capability 2.x/3.x)

  • l2_l1_read_hit_rate
    • Hit rate at L2 cache for all read requests from L1 cache
  • l2_tex_read_hit_rate
    • Hit rate at L2 cache for all read requests from texture cache

Maxwell/Pascal(Capability 5.x/6.x)

  • l2_tex_read_hit_rate
    • Hit rate at L2 cache for all read requests from texture cache

2.3 Transactions

L1 Cache

Global data

  • gld_transactions
    • Number of global memory load transactions
  • gld_transactions_per_request
    • Average number of global memory load transactions performed for each global memory load
  • gst_transactions
    • Number of global memory store transactions
  • gst_transactions_per_request
    • Average number of global memory store transactions performed for each global memory store

Local data

  • local_load_transactions
    • Number of local memory load transactions
  • local_load_transactions_per_request
    • Average number of local memory load transactions performed for each local memory load
  • local_store_transactions
    • Number of local memory store transactions
  • local_store_transactions_per_request
    • Average number of local memory store transactions performed for each local memory store

L2 Cache

Fermi/Kepler (Capability 2.x/3.x)

  • l2_l1_read_transactions
    • Memory read transactions seen at L2 cache for all read requests from L1 cache
  • l2_l1_write_transactions
    • Memory write transactions seen at L2 cache for all write requests from L1 cache

Maxwell/Pascal(Capability 5.x/6.x)

  • l2_tex_read_transactions
    • Memory read transactions seen at L2 cache for read requests from the texture cache
  • l2_tex_write_transactions

    Both

  • l2_read_transactions
    • Memory read transactions seen at L2 cache for all read requests
  • l2_write_transactions
    • Memory write transactions seen at L2 cache for all write requests

Only in Kepler

  • nc_l2_read_transactions
    • Memory read transactions seen at L2 cache for non coherent global read requests

备注

  • 自Kepler架构以来，L1 Cache 对 global data 的默认策略是 bypassing ，只有Fermi架构L1 Cache对 global data 是既可读又可写的，但是不能保持cache coherence 。
  • 那么为了保证 cache coherence,nvidia 采取了较为极端的做法，那就是bypassing L1 Cache ，并且在Maxwell 与 Pascal 架构中，与Tex Cache 合并，设置为 Read Only , 但我认为其效果并不佳。最新架构volta又将其架构改为 Fermi 中 L1 Cache 与 Shared memory 可配置的模式。
  • 可知，在Maxwell 与 Pascal 架构中，我们就将 tex cache 看成 L1 Data Cache

GDRAM

  • dram_read_transactions
    • Device memory read transactions
  • dram_write_transactions
    • Device memory write transactions

DRAM

  • sysmem_read_transactions
    • System memory read transactions
  • sysmem_write_transactions
    • System memory write transactions

​ Influence by L2 Hit Rate

Reference

Read more at: http://docs.nvidia.com/cuda/profiler-users-guide/index.html#ixzz4t4vGKod8
Follow us: @GPUComputing on Twitter | NVIDIA on Facebook