package xpu_timer.hook

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: KernelTrace

• repeated uint32 shapes = 1

  b s s h

Used in: HostTrace

• uint32 collected = 1

• uint32 uncollectable = 2

Used in: KernelTrace

• repeated GroupedMatmulDebugData.BMNK bmnks = 1

• string dtype = 2

• string api = 3

• double tflops = 4

Used in: GroupedMatmulDebugData

• repeated uint32 shapes = 1

  b m n k

Used in: KernelTraces

• uint64 start_us = 1

• uint32 dur_us = 2

• string name = 3

• uint32 count = 4

• oneof debug_data

  • GcDebugData gc_debug = 5

Used in: InterceptSymbolByOffset

• string func_name = 1

• string func_type = 2

• string operation = 3

• string algo = 4

• string dtype = 5

• string coll_type = 6

• bool only_trace = 7

• map<int64, InterceptSymbol> symbols = 1

Used in: KernelTraces

• uint32 kernel_type = 1

  we need configure which kind of kernel to dump trace the type is from brpc server to shm, so `kernel_type` needs to be primitive type

• uint32 trace_code = 2

• uint64 start_us = 3

• uint32 dur_us = 4

• uint32 delay_us = 5

• uint64 trace_id = 6

• oneof debug_data

  • MatmulDebugData mm_debug = 7

  • FaDebugData fa_debug = 8

  • NcclDebugData nccl_debug = 9

  • MemoryDebugData memory_debug = 10

  • GroupedMatmulDebugData grouped_mm_debug = 12

• bool is_host = 11

• repeated KernelTrace traces = 1

• repeated HostTrace host_traces = 2

• uint32 pid = 3

• uint32 rank = 4

Used in: KernelTrace, MatmulInfo

• repeated uint32 shapes = 1

  b m n k

• string dtype = 2

• string trans = 3

• repeated int64 lds = 4

  ld a b c

• repeated int64 strides = 5

  strides a b c

• int32 algo = 6

• string api = 7

Used in: MatmulInfos

• optional MatmulDebugData mm_debug = 1

• repeated double tflops = 2

Used in: RankMatmulInfo

• repeated MatmulInfo infos = 1

Used in: KernelTrace

• uint64 size = 1

  bytes

• string direction = 2

Used in: KernelTrace

• repeated uint32 grids = 1

  dim3

• repeated uint32 blocks = 2

  dim3

• uint64 comm_hash = 3

• uint64 input_size_in_bytes = 4

  diff coll operator are different

• string dtype = 5

• uint32 ranks = 6

• uint32 nodes = 7

• uint64 seq = 8

• uint64 problem_size = 9

• uint32 send_recv_type = 10

• map<int64, MatmulInfos> mm_infos = 1