Rocprofiler测试

Hi202402172024-09-21 21:35

Rocprofiler测试

Rocprofiler测试

一.参考链接

二.测试过程

1.登录服务器

bash 复制代码 
.TODO

2.使用smi获取列表

bash 复制代码 
rocm-smi

输出

bash 复制代码 
=========================================== ROCm System Management Interface ===========================================
===================================================== Concise Info =====================================================
Device  Node  IDs              Temp    Power     Partitions          SCLK    MCLK    Fan    Perf  PwrCap  VRAM%  GPU%
              (DID,     GUID)  (Edge)  (Socket)  (Mem, Compute, ID)
========================================================================================================================
0       1     0x66a1,   3820   35.0°C  20.0W     N/A, N/A, 0         860Mhz  350Mhz  9.41%  auto  190.0W  0%     0%
1       2     0x66a1,   22570  38.0°C  17.0W     N/A, N/A, 0         860Mhz  350Mhz  9.41%  auto  190.0W  0%     0%
========================================================================================================================
================================================= End of ROCm SMI Log ==================================================

3.使用rocminfo获取Agent信息

在 ROCm（Radeon Open Compute）平台中，Agent 通常指的是计算设备或处理单元，这些可以是 CPU 或 GPU。每个 Agent 可以执行计算任务并具有自己的计算资源，如计算核心、内存等。在 ROCm 的程序模型中，Agent 是负责执行特定任务的实体，当你使用 ROCm 进行并行计算时，任务通常会分配给不同的 Agent 来处理。Agent 是 ROCm 的异构计算环境中进行任务调度和管理的基本单元之一

bash 复制代码 
rocminfo

输出

bash 复制代码 
*******
Agent 2
*******
  Name:                    gfx906
  Uuid:                    GPU-021860c17348c2f7
  Marketing Name:          AMD Radeon (TM) Pro VII
  Vendor Name:             AMD
  Feature:                 KERNEL_DISPATCH
  Profile:                 BASE_PROFILE
  Float Round Mode:        NEAR
  Max Queue Number:        128(0x80)
  Queue Min Size:          64(0x40)
  Queue Max Size:          131072(0x20000)
  Queue Type:              MULTI
  Node:                    1
  Device Type:             GPU
  Cache Info:
    L1:                      16(0x10) KB
    L2:                      8192(0x2000) KB
  Chip ID:                 26273(0x66a1)
  ASIC Revision:           1(0x1)
  Cacheline Size:          64(0x40)
  Max Clock Freq. (MHz):   1700
  BDFID:                   1792
  Internal Node ID:        1
  Compute Unit:            60
  SIMDs per CU:            4
  Shader Engines:          4
  Shader Arrs. per Eng.:   1
  WatchPts on Addr. Ranges:4
  Coherent Host Access:    FALSE
  Memory Properties:
  Features:                KERNEL_DISPATCH
  Fast F16 Operation:      TRUE
  Wavefront Size:          64(0x40)
  Workgroup Max Size:      1024(0x400)
  Workgroup Max Size per Dimension:
    x                        1024(0x400)
    y                        1024(0x400)
    z                        1024(0x400)
  Max Waves Per CU:        40(0x28)
  Max Work-item Per CU:    2560(0xa00)
  Grid Max Size:           4294967295(0xffffffff)
  Grid Max Size per Dimension:
    x                        4294967295(0xffffffff)
    y                        4294967295(0xffffffff)
    z                        4294967295(0xffffffff)
  Max fbarriers/Workgrp:   32
  Packet Processor uCode:: 472
  SDMA engine uCode::      145
  IOMMU Support::          None
  Pool Info:
    Pool 1
      Segment:                 GLOBAL; FLAGS: COARSE GRAINED
      Size:                    16760832(0xffc000) KB
      Allocatable:             TRUE
      Alloc Granule:           4KB
      Alloc Recommended Granule:2048KB
      Alloc Alignment:         4KB
      Accessible by all:       FALSE
    Pool 2
      Segment:                 GLOBAL; FLAGS: EXTENDED FINE GRAINED
      Size:                    16760832(0xffc000) KB
      Allocatable:             TRUE
      Alloc Granule:           4KB
      Alloc Recommended Granule:2048KB
      Alloc Alignment:         4KB
      Accessible by all:       FALSE
    Pool 3
      Segment:                 GROUP
      Size:                    64(0x40) KB
      Allocatable:             FALSE
      Alloc Granule:           0KB
      Alloc Recommended Granule:0KB
      Alloc Alignment:         0KB
      Accessible by all:       FALSE
  ISA Info:
    ISA 1
      Name:                    amdgcn-amd-amdhsa--gfx906:sramecc+:xnack-
      Machine Models:          HSA_MACHINE_MODEL_LARGE
      Profiles:                HSA_PROFILE_BASE
      Default Rounding Mode:   NEAR
      Default Rounding Mode:   NEAR
      Fast f16:                TRUE
      Workgroup Max Size:      1024(0x400)
      Workgroup Max Size per Dimension:
        x                        1024(0x400)
        y                        1024(0x400)
        z                        1024(0x400)
      Grid Max Size:           4294967295(0xffffffff)
      Grid Max Size per Dimension:
        x                        4294967295(0xffffffff)
        y                        4294967295(0xffffffff)
        z                        4294967295(0xffffffff)
      FBarrier Max Size:       32
*******

4.准备测试用例

bash 复制代码 
tee ROCmMatrixTranspose.cpp<<-'EOF'
#include <iostream>
// hip header file
#include <hip/hip_runtime.h>
// roctx header file
#include <roctracer/roctx.h>

#define WIDTH 1024
#define NUM (WIDTH * WIDTH)
#define THREADS_PER_BLOCK_X 4
#define THREADS_PER_BLOCK_Y 4
#define THREADS_PER_BLOCK_Z 1

// Device (Kernel) function, it must be void
__global__ void matrixTranspose(float* out, float* in, const int width) {
  int x = hipBlockDim_x * hipBlockIdx_x + hipThreadIdx_x;
  int y = hipBlockDim_y * hipBlockIdx_y + hipThreadIdx_y;
  out[y * width + x] = in[x * width + y];
}

// CPU implementation of matrix transpose
void matrixTransposeCPUReference(float* output, float* input, const unsigned int width) {
  for (unsigned int j = 0; j < width; j++) {
    for (unsigned int i = 0; i < width; i++) {
      output[i * width + j] = input[j * width + i];
    }
  }
}

int main() {
  float* Matrix;
  float* TransposeMatrix;
  float* cpuTransposeMatrix;

  float* gpuMatrix;
  float* gpuTransposeMatrix;

  hipDeviceProp_t devProp;
  hipGetDeviceProperties(&devProp, 0);

  std::cout << "Device name " << devProp.name << std::endl;

  int i;
  int errors;

  Matrix = (float*)malloc(NUM * sizeof(float));
  TransposeMatrix = (float*)malloc(NUM * sizeof(float));
  cpuTransposeMatrix = (float*)malloc(NUM * sizeof(float));

  // initialize the input data
  for (i = 0; i < NUM; i++) {
    Matrix[i] = (float)i * 10.0f;
  }

  // allocate the memory on the device side
  hipMalloc((void**)&gpuMatrix, NUM * sizeof(float));
  hipMalloc((void**)&gpuTransposeMatrix, NUM * sizeof(float));

  uint32_t iterations = 1;
  while (iterations-- > 0) {
    std::cout << "## Iteration (" << iterations << ") #################" << std::endl;

    // Memory transfer from host to device
    hipMemcpy(gpuMatrix, Matrix, NUM * sizeof(float), hipMemcpyHostToDevice);

    roctxMark("ROCTX-MARK: before hipLaunchKernel");
    roctxRangePush("ROCTX-RANGE: hipLaunchKernel");

    roctx_range_id_t roctx_id = roctxRangeStartA("roctx_range with id");

    // Lauching kernel from host
    hipLaunchKernelGGL(
        matrixTranspose, dim3(WIDTH / THREADS_PER_BLOCK_X, WIDTH / THREADS_PER_BLOCK_Y),
        dim3(THREADS_PER_BLOCK_X, THREADS_PER_BLOCK_Y), 0, 0, gpuTransposeMatrix, gpuMatrix, WIDTH);

    roctxRangeStop(roctx_id);
    roctxMark("ROCTX-MARK: after hipLaunchKernel");

    // Memory transfer from device to host
    roctxRangePush("ROCTX-RANGE: hipMemcpy");

    hipMemcpy(TransposeMatrix, gpuTransposeMatrix, NUM * sizeof(float), hipMemcpyDeviceToHost);

    roctxRangePop();  // for "hipMemcpy"
    roctxRangePop();  // for "hipLaunchKernel"

    // CPU MatrixTranspose computation
    matrixTransposeCPUReference(cpuTransposeMatrix, Matrix, WIDTH);

    // verify the results
    errors = 0;
    double eps = 1.0E-6;
    for (i = 0; i < NUM; i++) {
      if (std::abs(TransposeMatrix[i] - cpuTransposeMatrix[i]) > eps) {
        errors++;
      }
    }
    if (errors != 0) {
      printf("FAILED: %d errors\n", errors);
    } else {
      printf("PASSED!\n");
    }
  }

  // free the resources on device side
  hipFree(gpuMatrix);
  hipFree(gpuTransposeMatrix);

  // free the resources on host side
  free(Matrix);
  free(TransposeMatrix);
  free(cpuTransposeMatrix);

  return errors;
}

EOF

/opt/rocm/bin/hipcc -c ROCmMatrixTranspose.cpp -o ROCmMatrixTranspose.cpp.o
/opt/rocm/bin/hipcc ROCmMatrixTranspose.cpp.o -o ROCmMatrixTranspose \
    /opt/rocm/lib/libamd_comgr.so.2.8.60200 /usr/lib/x86_64-linux-gnu/libnuma.so /opt/rocm/lib/libroctx64.so	
./ROCmMatrixTranspose

5.The hardware counters are called the basic counters

bash 复制代码 
rocprof --list-basic | grep -A 2  "gpu-agent2"

输出

bash 复制代码 
  gpu-agent2 : TCC_EA1_WRREQ[0-15] : Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA1_WRREQ_64B[0-15] : Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA1_WRREQ_STALL[0-15] : Number of cycles a write request was stalled.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA1_RDREQ[0-15] : Number of TCC/EA read requests (either 32-byte or 64-byte)
      block TCC has 4 counters

  gpu-agent2 : TCC_EA1_RDREQ_32B[0-15] : Number of 32-byte TCC/EA read requests
      block TCC has 4 counters

  gpu-agent2 : GRBM_COUNT : Tie High - Count Number of Clocks
      block GRBM has 2 counters

  gpu-agent2 : GRBM_GUI_ACTIVE : The GUI is Active
      block GRBM has 2 counters

  gpu-agent2 : SQ_WAVES : Count number of waves sent to SQs. (per-simd, emulated, global)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_VALU : Number of VALU instructions issued. (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_VMEM_WR : Number of VMEM write instructions issued (including FLAT). (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_VMEM_RD : Number of VMEM read instructions issued (including FLAT). (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_SALU : Number of SALU instructions issued. (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_SMEM : Number of SMEM instructions issued. (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_FLAT : Number of FLAT instructions issued. (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_FLAT_LDS_ONLY : Number of FLAT instructions issued that read/wrote only from/to LDS (only works if EARLY_TA_DONE is enabled). (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_LDS : Number of LDS instructions issued (including FLAT). (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_INSTS_GDS : Number of GDS instructions issued. (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_WAIT_INST_LDS : Number of wave-cycles spent waiting for LDS instruction issue. In units of 4 cycles. (per-simd, nondeterministic)
      block SQ has 8 counters

  gpu-agent2 : SQ_ACTIVE_INST_VALU : regspec 71? Number of cycles the SQ instruction arbiter is working on a VALU instruction. (per-simd, nondeterministic). Units in quad-cycles(4 cycles)
      block SQ has 8 counters

  gpu-agent2 : SQ_INST_CYCLES_SALU : Number of cycles needed to execute non-memory read scalar operations. (per-simd, emulated)
      block SQ has 8 counters

  gpu-agent2 : SQ_THREAD_CYCLES_VALU : Number of thread-cycles used to execute VALU operations (similar to INST_CYCLES_VALU but multiplied by # of active threads). (per-simd)
      block SQ has 8 counters

  gpu-agent2 : SQ_LDS_BANK_CONFLICT : Number of cycles LDS is stalled by bank conflicts. (emulated)
      block SQ has 8 counters

  gpu-agent2 : TA_TA_BUSY[0-15] : TA block is busy. Perf_Windowing not supported for this counter.
      block TA has 2 counters

  gpu-agent2 : TA_FLAT_READ_WAVEFRONTS[0-15] : Number of flat opcode reads processed by the TA.
      block TA has 2 counters

  gpu-agent2 : TA_FLAT_WRITE_WAVEFRONTS[0-15] : Number of flat opcode writes processed by the TA.
      block TA has 2 counters

  gpu-agent2 : TCC_HIT[0-15] : Number of cache hits.
      block TCC has 4 counters

  gpu-agent2 : TCC_MISS[0-15] : Number of cache misses. UC reads count as misses.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA_WRREQ[0-15] : Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Atomics may travel over the same interface and are generally classified as write requests. This does not include probe commands.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA_WRREQ_64B[0-15] : Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA_WRREQ_STALL[0-15] : Number of cycles a write request was stalled.
      block TCC has 4 counters

  gpu-agent2 : TCC_EA_RDREQ[0-15] : Number of TCC/EA read requests (either 32-byte or 64-byte)
      block TCC has 4 counters

  gpu-agent2 : TCC_EA_RDREQ_32B[0-15] : Number of 32-byte TCC/EA read requests
      block TCC has 4 counters

  gpu-agent2 : TCP_TCP_TA_DATA_STALL_CYCLES[0-15] : TCP stalls TA data interface. Now Windowed.
      block TCP has 4 counters

6.The derived metrics are defined on top of the basic counters using mathematical expression

bash 复制代码 
rocprof --list-derived | grep -A 2  "gpu-agent2"

输出

bash 复制代码 
  gpu-agent2 : TCC_EA1_RDREQ_32B_sum : Number of 32-byte TCC/EA read requests. Sum over TCC EA1s.
      TCC_EA1_RDREQ_32B_sum = sum(TCC_EA1_RDREQ_32B,16)

  gpu-agent2 : TCC_EA1_RDREQ_sum : Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC EA1s.
      TCC_EA1_RDREQ_sum = sum(TCC_EA1_RDREQ,16)

  gpu-agent2 : TCC_EA1_WRREQ_sum : Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC EA1s.
      TCC_EA1_WRREQ_sum = sum(TCC_EA1_WRREQ,16)

  gpu-agent2 : TCC_EA1_WRREQ_64B_sum : Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC EA1s.
      TCC_EA1_WRREQ_64B_sum = sum(TCC_EA1_WRREQ_64B,16)

  gpu-agent2 : TCC_WRREQ1_STALL_max : Number of cycles a write request was stalled. Max over TCC instances.
      TCC_WRREQ1_STALL_max = max(TCC_EA1_WRREQ_STALL,16)

  gpu-agent2 : RDATA1_SIZE : The total kilobytes fetched from the video memory. This is measured on EA1s.
      RDATA1_SIZE = (TCC_EA1_RDREQ_32B_sum*32+(TCC_EA1_RDREQ_sum-TCC_EA1_RDREQ_32B_sum)*64)

  gpu-agent2 : WDATA1_SIZE : The total kilobytes written to the video memory. This is measured on EA1s.
      WDATA1_SIZE = ((TCC_EA1_WRREQ_sum-TCC_EA1_WRREQ_64B_sum)*32+TCC_EA1_WRREQ_64B_sum*64)

  gpu-agent2 : FETCH_SIZE : The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account.
      FETCH_SIZE = (TCC_EA_RDREQ_32B_sum*32+(TCC_EA_RDREQ_sum-TCC_EA_RDREQ_32B_sum)*64+RDATA1_SIZE)/1024

  gpu-agent2 : WRITE_SIZE : The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account.
      WRITE_SIZE = ((TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)*32+TCC_EA_WRREQ_64B_sum*64+WDATA1_SIZE)/1024

  gpu-agent2 : WRITE_REQ_32B : The total number of 32-byte effective memory writes.
      WRITE_REQ_32B = (TCC_EA_WRREQ_sum-TCC_EA_WRREQ_64B_sum)+(TCC_EA1_WRREQ_sum-TCC_EA1_WRREQ_64B_sum)+(TCC_EA_WRREQ_64B_sum+TCC_EA1_WRREQ_64B_sum)*2

  gpu-agent2 : TA_BUSY_avr : TA block is busy. Average over TA instances.
      TA_BUSY_avr = avr(TA_TA_BUSY,16)

  gpu-agent2 : TA_BUSY_max : TA block is busy. Max over TA instances.
      TA_BUSY_max = max(TA_TA_BUSY,16)

  gpu-agent2 : TA_BUSY_min : TA block is busy. Min over TA instances.
      TA_BUSY_min = min(TA_TA_BUSY,16)

  gpu-agent2 : TA_FLAT_READ_WAVEFRONTS_sum : Number of flat opcode reads processed by the TA. Sum over TA instances.
      TA_FLAT_READ_WAVEFRONTS_sum = sum(TA_FLAT_READ_WAVEFRONTS,16)

  gpu-agent2 : TA_FLAT_WRITE_WAVEFRONTS_sum : Number of flat opcode writes processed by the TA. Sum over TA instances.
      TA_FLAT_WRITE_WAVEFRONTS_sum = sum(TA_FLAT_WRITE_WAVEFRONTS,16)

  gpu-agent2 : TCC_HIT_sum : Number of cache hits. Sum over TCC instances.
      TCC_HIT_sum = sum(TCC_HIT,16)

  gpu-agent2 : TCC_MISS_sum : Number of cache misses. Sum over TCC instances.
      TCC_MISS_sum = sum(TCC_MISS,16)

  gpu-agent2 : TCC_EA_RDREQ_32B_sum : Number of 32-byte TCC/EA read requests. Sum over TCC instances.
      TCC_EA_RDREQ_32B_sum = sum(TCC_EA_RDREQ_32B,16)

  gpu-agent2 : TCC_EA_RDREQ_sum : Number of TCC/EA read requests (either 32-byte or 64-byte). Sum over TCC instances.
      TCC_EA_RDREQ_sum = sum(TCC_EA_RDREQ,16)

  gpu-agent2 : TCC_EA_WRREQ_sum : Number of transactions (either 32-byte or 64-byte) going over the TC_EA_wrreq interface. Sum over TCC instances.
      TCC_EA_WRREQ_sum = sum(TCC_EA_WRREQ,16)

  gpu-agent2 : TCC_EA_WRREQ_64B_sum : Number of 64-byte transactions going (64-byte write or CMPSWAP) over the TC_EA_wrreq interface. Sum over TCC instances.
      TCC_EA_WRREQ_64B_sum = sum(TCC_EA_WRREQ_64B,16)

  gpu-agent2 : TCC_WRREQ_STALL_max : Number of cycles a write request was stalled. Max over TCC instances.
      TCC_WRREQ_STALL_max = max(TCC_EA_WRREQ_STALL,16)

  gpu-agent2 : TCP_TCP_TA_DATA_STALL_CYCLES_sum : Total number of TCP stalls TA data interface.
      TCP_TCP_TA_DATA_STALL_CYCLES_sum = sum(TCP_TCP_TA_DATA_STALL_CYCLES,16)

  gpu-agent2 : TCP_TCP_TA_DATA_STALL_CYCLES_max : Maximum number of TCP stalls TA data interface.
      TCP_TCP_TA_DATA_STALL_CYCLES_max = max(TCP_TCP_TA_DATA_STALL_CYCLES,16)

  gpu-agent2 : VFetchInsts : The average number of vector fetch instructions from the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that fetch from video memory.
      VFetchInsts = (SQ_INSTS_VMEM_RD-TA_FLAT_READ_WAVEFRONTS_sum)/SQ_WAVES

  gpu-agent2 : VWriteInsts : The average number of vector write instructions to the video memory executed per work-item (affected by flow control). Excludes FLAT instructions that write to video memory.
      VWriteInsts = (SQ_INSTS_VMEM_WR-TA_FLAT_WRITE_WAVEFRONTS_sum)/SQ_WAVES

  gpu-agent2 : FlatVMemInsts : The average number of FLAT instructions that read from or write to the video memory executed per work item (affected by flow control). Includes FLAT instructions that read from or write to scratch.
      FlatVMemInsts = (SQ_INSTS_FLAT-SQ_INSTS_FLAT_LDS_ONLY)/SQ_WAVES

  gpu-agent2 : LDSInsts : The average number of LDS read or LDS write instructions executed per work item (affected by flow control).  Excludes FLAT instructions that read from or write to LDS.
      LDSInsts = (SQ_INSTS_LDS-SQ_INSTS_FLAT_LDS_ONLY)/SQ_WAVES

  gpu-agent2 : FlatLDSInsts : The average number of FLAT instructions that read or write to LDS executed per work item (affected by flow control).
      FlatLDSInsts = SQ_INSTS_FLAT_LDS_ONLY/SQ_WAVES

  gpu-agent2 : VALUUtilization : The percentage of active vector ALU threads in a wave. A lower number can mean either more thread divergence in a wave or that the work-group size is not a multiple of 64. Value range: 0% (bad), 100% (ideal - no thread divergence).
      VALUUtilization = 100*SQ_THREAD_CYCLES_VALU/(SQ_ACTIVE_INST_VALU*MAX_WAVE_SIZE)

  gpu-agent2 : VALUBusy : The percentage of GPUTime vector ALU instructions are processed. Value range: 0% (bad) to 100% (optimal).
      VALUBusy = 100*SQ_ACTIVE_INST_VALU*4/SIMD_NUM/GRBM_GUI_ACTIVE

  gpu-agent2 : SALUBusy : The percentage of GPUTime scalar ALU instructions are processed. Value range: 0% (bad) to 100% (optimal).
      SALUBusy = 100*SQ_INST_CYCLES_SALU*4/SIMD_NUM/GRBM_GUI_ACTIVE

  gpu-agent2 : FetchSize : The total kilobytes fetched from the video memory. This is measured with all extra fetches and any cache or memory effects taken into account.
      FetchSize = FETCH_SIZE

  gpu-agent2 : WriteSize : The total kilobytes written to the video memory. This is measured with all extra fetches and any cache or memory effects taken into account.
      WriteSize = WRITE_SIZE

  gpu-agent2 : MemWrites32B : The total number of effective 32B write transactions to the memory
      MemWrites32B = WRITE_REQ_32B

  gpu-agent2 : L2CacheHit : The percentage of fetch, write, atomic, and other instructions that hit the data in L2 cache. Value range: 0% (no hit) to 100% (optimal).
      L2CacheHit = 100*sum(TCC_HIT,16)/(sum(TCC_HIT,16)+sum(TCC_MISS,16))

  gpu-agent2 : MemUnitStalled : The percentage of GPUTime the memory unit is stalled. Try reducing the number or size of fetches and writes if possible. Value range: 0% (optimal) to 100% (bad).
      MemUnitStalled = 100*max(TCP_TCP_TA_DATA_STALL_CYCLES,16)/GRBM_GUI_ACTIVE/SE_NUM

  gpu-agent2 : WriteUnitStalled : The percentage of GPUTime the Write unit is stalled. Value range: 0% to 100% (bad).
      WriteUnitStalled = 100*TCC_WRREQ_STALL_max/GRBM_GUI_ACTIVE

  gpu-agent2 : LDSBankConflict : The percentage of GPUTime LDS is stalled by bank conflicts. Value range: 0% (optimal) to 100% (bad).
      LDSBankConflict = 100*SQ_LDS_BANK_CONFLICT/GRBM_GUI_ACTIVE/CU_NUM

  gpu-agent2 : GPUBusy : The percentage of time GPU was busy.
      GPUBusy = 100*GRBM_GUI_ACTIVE/GRBM_COUNT

  gpu-agent2 : Wavefronts : Total wavefronts.
      Wavefronts = SQ_WAVES

  gpu-agent2 : VALUInsts : The average number of vector ALU instructions executed per work-item (affected by flow control).
      VALUInsts = SQ_INSTS_VALU/SQ_WAVES

  gpu-agent2 : SALUInsts : The average number of scalar ALU instructions executed per work-item (affected by flow control).
      SALUInsts = SQ_INSTS_SALU/SQ_WAVES

  gpu-agent2 : SFetchInsts : The average number of scalar fetch instructions from the video memory executed per work-item (affected by flow control).
      SFetchInsts = SQ_INSTS_SMEM/SQ_WAVES

  gpu-agent2 : GDSInsts : The average number of GDS read or GDS write instructions executed per work item (affected by flow control).
      GDSInsts = SQ_INSTS_GDS/SQ_WAVES

  gpu-agent2 : MemUnitBusy : The percentage of GPUTime the memory unit is active. The result includes the stall time (MemUnitStalled). This is measured with all extra fetches and writes and any cache or memory effects taken into account. Value range: 0% to 100% (fetch-bound).
      MemUnitBusy = 100*max(TA_TA_BUSY,16)/GRBM_GUI_ACTIVE/SE_NUM

  gpu-agent2 : ALUStalledByLDS : The percentage of GPUTime ALU units are stalled by the LDS input queue being full or the output queue being not ready. If there are LDS bank conflicts, reduce them. Otherwise, try reducing the number of LDS accesses if possible. Value range: 0% (optimal) to 100% (bad).
      ALUStalledByLDS = 100*SQ_WAIT_INST_LDS*4/SQ_WAVES/GRBM_GUI_ACTIVE

7.Profing

bash 复制代码 
tee input.txt<<-'EOF'
pmc : Wavefronts, VALUInsts, SALUInsts, SFetchInsts,FlatVMemInsts,
LDSInsts, FlatLDSInsts, GDSInsts, VALUUtilization, FetchSize,
WriteSize, L2CacheHit, VWriteInsts, GPUBusy, VALUBusy, SALUBusy,
MemUnitStalled, WriteUnitStalled, LDSBankConflict, MemUnitBusy
# Filter by dispatches range, GPU index and kernel names
# supported range formats: "3:9", "3:", "3"
range: 0 : 1
gpu: 0
kernel:matrixTranspose
EOF

rocprof -i input.txt ./ROCmMatrixTranspose
cat /root/input.csv
rocprofv2 -i input.txt ./ROCmMatrixTranspose
rocprofv2 --hsa-trace ./ROCmMatrixTranspose

输出

bash 复制代码 
RPL: on '240920_102257' from '/opt/rocm-6.2.0' in '/root'
RPL: profiling '"./ROCmMatrixTranspose"'
RPL: input file 'input.txt'
RPL: output dir '/tmp/rpl_data_240920_102257_47892'

RPL: result dir '/tmp/rpl_data_240920_102257_47892/input0_results_240920_102257'
ROCProfiler: input from "/tmp/rpl_data_240920_102257_47892/input0.xml"
  gpu_index = 0
  kernel = matrixTranspose
  range = 0:1
  4 metrics
    Wavefronts, VALUInsts, SALUInsts, SFetchInsts
Device name AMD Radeon (TM) Pro VII
## Iteration (0) #################
PASSED!

ROCPRofiler: 1 contexts collected, output directory /tmp/rpl_data_240920_102257_47892/input0_results_240920_102257
File '/root/input.csv' is generating
Index,KernelName,gpu-id,queue-id,queue-index,pid,tid,grd,wgr,lds,scr,arch_vgpr,accum_vgpr,sgpr,wave_size,sig,obj,Wavefronts,VALUInsts,SALUInsts,SFetchInsts
0,"matrixTranspose(float*, float*, int) [clone .kd]",1,0,0,48178,48178,1048576,16,0,0,8,0,16,64,0x0,0x742031870880,65536.0000000000,14.0000000000,4.0000000000,3.0000000000

ROCProfilerV2: Collecting the following counters:
- Wavefronts
- VALUInsts
- SALUInsts
- SFetchInsts
Enabling Counter Collection
Device name AMD Radeon (TM) Pro VII
## Iteration (0) #################
PASSED!
Dispatch_ID(0), GPU_ID(1), Queue_ID(1), Process_ID(48209), Thread_ID(48209), Grid_Size(1048576), Workgroup_Size(16), LDS_Per_Workgroup(0), Scratch_Per_Workitem(0), Arch_VGPR(8), Accum_VGPR(0), SGPR(16), Wave_Size(64), Kernel_Name("matrixTranspose(float*, float*, int) (.kd)"), Begin_Timestamp(951172884265490), End_Timestamp(951172884454463), Correlation_ID(0), SALUInsts(4.000000), SFetchInsts(3.000000), VALUInsts(14.000000), Wavefronts(65536.000000)
相关推荐
探索云原生1 天前
大模型推理指南：使用 vLLM 实现高效推理
ai·云原生·kubernetes·gpu·vllm
若石之上4 天前
DeepSpeed：PyTorch优化库，使模型分布式训练能高效使用内存和更快速
pytorch·内存·gpu·deepspeed·速度·zero
luoganttcc5 天前
ubuntu.24安装cuda
cuda
qiang425 天前
想租用显卡训练自己的网络？AutoDL保姆级使用教程（PyCharm版）
pycharm·gpu·autodl·租显卡
扫地的小何尚8 天前
NVIDIA RTX 系统上使用 llama.cpp 加速 LLM
人工智能·aigc·llama·gpu·nvidia·cuda·英伟达
藓类少女8 天前
【深度学习】使用硬件加速模型训练速度
人工智能·深度学习·分布式训练·gpu
centurysee10 天前
【一文搞懂】GPU硬件拓扑与传输速度
gpu·nvidia
吃肉夹馍不要夹馍10 天前
CublasLt 极简入门
cuda·cublas·gemm·cublaslt
Code-world-112 天前
Ubuntu系统安装NVIDIA驱动、CUDA、PyTorch等GPU深度学习环境
linux·pytorch·深度学习·cuda·深度强化学习
热门推荐
01(欧拉)openEuler系统添加网卡文件配置流程、(欧拉)openEuler系统手动配置ipv6地址流程、(欧拉)openEuler系统网络管理说明02PyTorch机器学习实现液态神经网络03【HarmonyOS】HUAWEI DevEco Studio 下载地址汇总04玄机平台应急响应—webshell查杀05Coze扣子平台完整体验和实践（附国内和国际版对比）06RAG 实践- Ollama+RagFlow 部署本地知识库07【目标跟踪】相机运动补偿08Unity中PICO实现 隔空取物 和 接触抓取物体09Ubuntu 20.04使用Livox mid 360 测试 FAST_LIO10量化方法怎么选？如何评估量化后的大模型LLM？