RTX3060 FP64测试与猜想

RTX3060 FP64测试与猜想

RTX3060 FP64测试与猜想

一.小结

  • RTX3060 compute capability为8.6,每个SM有2个FP64 core。每个cycle可输出2个fp64的结果

  • RTX3060 有4个subcore,这2个core怎么给4个sub_core分呢

  • 执行FP64 DADD指令时,MIO PQ利用率超20%(FADD指令不存在该现象),且fp64 pipe的利用率最多为84%

  • 每个smsp 执行一条DADD warp指令 pipe_fp64_cycles_active 增加16个cycle,4个smsp一起运行一条DADD warp指令仍是16个cycle

  • 猜测:

    • smsp按 1DADD/cycle 交替发送给2个FP64 core,一个warp需要16个cycle(32inst/16cycle->2inst/cycle)
    • 如果4个sub core同时按这个速度发,则超过了FP64的处理能力(8inst/cycle > 2inst/cycle),但pipe_fp64_cycles_active没有增加
    • 说明,在发射FP64指令之前会检测资源的可用性,如果不足,则不发射,pipe_fp64_cycles_active也就不会增加
    • 也就解释了4个sub core一起执行时,pipe_fp64_cycles_active.max还是16个cycle
    • 执行FP64指令时,4个subcore通过MIO共享FP64实际的执行单元
  • How does four subcores in a single SM share two FP64 cores?

二.查看FP64的峰值性能

c 复制代码
tee fp64_peak_sustained.cu<<-'EOF'
#include <cuda_runtime.h>
#include <cuda.h>
__global__ void fake_kernel(){}
int main(int argc,char *argv[])
{
    fake_kernel<<<1, 1>>>();cudaDeviceSynchronize();
}
EOF
/usr/local/cuda/bin/nvcc -std=c++17 -arch=sm_86 -lineinfo  -o fp64_peak_sustained fp64_peak_sustained.cu \
    -I /usr/local/cuda/include -L /usr/local/cuda/lib64 -lcuda
/usr/local/NVIDIA-Nsight-Compute/ncu --metrics \
sm__sass_thread_inst_executed_op_fp64_pred_on.avg.peak_sustained,\
sm__sass_thread_inst_executed_op_fp64_pred_on.sum.peak_sustained ./fp64_peak_sustained

输出

bash 复制代码
fake_kernel() (1, 1, 1)x(1, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
---------------------------------------------------------------- ----------- ------------
Metric Name                                                      Metric Unit Metric Value
---------------------------------------------------------------- ----------- ------------
sm__sass_thread_inst_executed_op_fp64_pred_on.avg.peak_sustained  inst/cycle            2  #每个sm的峰值性能
sm__sass_thread_inst_executed_op_fp64_pred_on.sum.peak_sustained  inst/cycle           56  #28个sm
---------------------------------------------------------------- ----------- ------------

三.打满FP64、FP32的利用率,对比差异

c 复制代码
tee fp64_test.cu<<-'EOF'
#include <iostream>
#include <cuda_runtime.h>
#include <iostream>
#include <vector>
#include <stdio.h>
#include <assert.h>
#include <cstdio>
#include <cuda.h>

__global__ void kernel_add_float(volatile float *input,volatile float *output)
{
    unsigned int tid  = threadIdx.x + blockIdx.x * blockDim.x;
    float l=input[tid];
    float r=output[tid];
    for(int i=0;i<256;i++)
    {
        l-=r;
    } 
    input[tid]=l;
}
__global__ void kernel_add_double(volatile double *input,volatile double *output)
{
    unsigned int tid  = threadIdx.x + blockIdx.x * blockDim.x;
    double left=input[tid];
    double right=output[tid];
    for(int i=0;i<256;i++)
    {
        left+=right;
    }       
    output[tid]=left;
}
EOF

/usr/local/cuda/bin/nvcc -std=c++17 -dc -lineinfo -arch=sm_86 -ptx fp64_test.cu -o fp64_test.ptx
/usr/local/cuda/bin/nvcc -arch=sm_86 fp64_test.ptx -cubin -o fp64_test.cubin
/usr/local/cuda/bin/nvcc -arch=sm_86 fp64_test.cubin -fatbin -o fp64_test.fatbin
cat fp64_test.ptx
/usr/local/cuda/bin/cuobjdump --dump-sass fp64_test.fatbin

# 删掉除DADD、FADD以外的指令
cuasm.py fp64_test.cubin fp64_test.cuasm
sed '/MOV/d' -i fp64_test.cuasm
sed '/S2R/d' -i fp64_test.cuasm
sed '/ULDC/d' -i fp64_test.cuasm
sed '/IMAD/d' -i fp64_test.cuasm
sed '/LDG/d' -i fp64_test.cuasm
sed '/STG/d' -i fp64_test.cuasm
sed '/F2F/d' -i fp64_test.cuasm

cuasm.py fp64_test.cuasm
/usr/local/cuda/bin/nvcc -arch=sm_86 fp64_test.cubin -fatbin -o fp64_test.fatbin
/usr/local/cuda/bin/cuobjdump --dump-sass fp64_test.fatbin
/usr/local/cuda/bin/cuobjdump --dump-resource-usage fp64_test.fatbin

tee fp64_test_main.cpp<<-'EOF'
#include <stdio.h>
#include <string.h>
#include <cuda_runtime.h>
#include <cuda.h>

int main(int argc,char *argv[])
{
    CUresult error;
    CUdevice cuDevice;
    cuInit(0);
    int deviceCount = 0;
    error = cuDeviceGetCount(&deviceCount);
    error = cuDeviceGet(&cuDevice, 0);
    if(error!=CUDA_SUCCESS)
        {
        printf("Error happened in get device!\n");
    }
    CUcontext cuContext;
    error = cuCtxCreate(&cuContext, 0, cuDevice);
    if(error!=CUDA_SUCCESS)
        {
        printf("Error happened in create context!\n");
    }
    int block_count=28*1000;int block_size=32*4*4;
    int thread_size=block_count*block_size;

    int data_size=sizeof(double)*thread_size;

    double *output_ptr=nullptr;
    double *input_ptr=nullptr;
    int cudaStatus=0;
    cudaStatus = cudaMalloc((void**)&input_ptr, data_size);
    cudaStatus = cudaMalloc((void**)&output_ptr, data_size);
    void *kernelParams[]= {(void*)&output_ptr, (void*)&input_ptr};

    CUmodule module;
    CUfunction double_function;
    CUfunction float_function;
    const char* module_file = "fp64_test.fatbin";
    const char* double_kernel_name = "_Z17kernel_add_doublePVdS0_";
    const char* float_kernel_name = "_Z16kernel_add_floatPVfS0_";
    
    error = cuModuleLoad(&module, module_file);
    if(error!=CUDA_SUCCESS)
        {
        printf("Error happened in load moudle %d!\n",error);
    }
    error = cuModuleGetFunction(&double_function, module, double_kernel_name);
    if(error!=CUDA_SUCCESS)
    {
        printf("get double_function error!\n");
    }
    error = cuModuleGetFunction(&float_function, module, float_kernel_name);
    if(error!=CUDA_SUCCESS)
    {
        printf("get float_kernel_name error!\n");
    }    
    cuLaunchKernel(double_function,
                    block_count, 1, 1,
                    block_size, 1, 1,
                    0,0,kernelParams, 0);
    cuLaunchKernel(float_function,
                    block_count, 1, 1,
                    block_size, 1, 1,
                    0,0,kernelParams, 0);
    cudaFree(output_ptr);
    cudaFree(input_ptr);
    cuModuleUnload(module);
    cuCtxDestroy(cuContext);
    return 0;
}
EOF
g++ fp64_test_main.cpp -o fp64_test_main -I /usr/local/cuda/include -L /usr/local/cuda/lib64 -lcudart -lcuda
/usr/local/NVIDIA-Nsight-Compute/ncu --metrics \
sm__inst_executed.avg.pct_of_peak_sustained_elapsed,\
smsp__inst_issued.sum,\
sm__issue_active.avg.pct_of_peak_sustained_elapsed,\
sm__pipe_fma_cycles_active.avg.pct_of_peak_sustained_elapsed,\
sm__pipe_fmaheavy_cycles_active.avg.pct_of_peak_sustained_elapsed,\
sm__inst_executed_pipe_cbu_pred_on_any.avg.pct_of_peak_sustained_elapsed,\
sm__mio_inst_issued.avg.pct_of_peak_sustained_elapsed,\
sm__mio_pq_read_cycles_active.avg.pct_of_peak_sustained_elapsed,\
sm__mio_pq_write_cycles_active.avg.pct_of_peak_sustained_elapsed,\
sm__mio_pq_write_cycles_active_pipe_lsu.avg.pct_of_peak_sustained_elapsed,\
sm__mio_pq_write_cycles_active_pipe_tex.avg.pct_of_peak_sustained_elapsed,\
sm__mioc_inst_issued.avg.pct_of_peak_sustained_elapsed,\
sm__mio_inst_issued.avg.pct_of_peak_sustained_elapsed,\
sm__pipe_fp64_cycles_active.avg.pct_of_peak_sustained_elapsed ./fp64_test_main

输出

bash 复制代码
kernel_add_double(volatile double *, volatile double *) (28000, 1, 1)x(512, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
------------------------------------------------------------------------- ----------- ------------
Metric Name                                                               Metric Unit Metric Value
------------------------------------------------------------------------- ----------- ------------
sm__inst_executed.avg.pct_of_peak_sustained_elapsed                                 %         1.32
sm__inst_executed_pipe_cbu_pred_on_any.avg.pct_of_peak_sustained_elapsed            %         0.02
sm__issue_active.avg.pct_of_peak_sustained_elapsed                                  %         1.32
sm__mio_inst_issued.avg.pct_of_peak_sustained_elapsed                               %         3.51
sm__mio_pq_read_cycles_active.avg.pct_of_peak_sustained_elapsed                     %            0
sm__mio_pq_write_cycles_active.avg.pct_of_peak_sustained_elapsed                    %        21.05
sm__mio_pq_write_cycles_active_pipe_lsu.avg.pct_of_peak_sustained_elapsed           %            0
sm__mio_pq_write_cycles_active_pipe_tex.avg.pct_of_peak_sustained_elapsed           %        21.05 # of cycles where register operands from the register file were
                                                                                                     written to MIO PQ, for the tex pipe
sm__mioc_inst_issued.avg.pct_of_peak_sustained_elapsed                              %         1.32
sm__pipe_fma_cycles_active.avg.pct_of_peak_sustained_elapsed                        %            0
sm__pipe_fmaheavy_cycles_active.avg.pct_of_peak_sustained_elapsed                   %            0
sm__pipe_fp64_cycles_active.avg.pct_of_peak_sustained_elapsed                       %        84.21  #利用率打不满
smsp__inst_issued.sum                                                            inst  115,136,000  #跟fp32相同的指令条数
------------------------------------------------------------------------- ----------- ------------

kernel_add_float(volatile float *, volatile float *) (28000, 1, 1)x(512, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
------------------------------------------------------------------------- ----------- ------------
Metric Name                                                               Metric Unit Metric Value
------------------------------------------------------------------------- ----------- ------------
sm__inst_executed.avg.pct_of_peak_sustained_elapsed                                 %        99.76
sm__inst_executed_pipe_cbu_pred_on_any.avg.pct_of_peak_sustained_elapsed            %         1.55
sm__issue_active.avg.pct_of_peak_sustained_elapsed                                  %        99.76
sm__mio_inst_issued.avg.pct_of_peak_sustained_elapsed                               %            0
sm__mio_pq_read_cycles_active.avg.pct_of_peak_sustained_elapsed                     %            0
sm__mio_pq_write_cycles_active.avg.pct_of_peak_sustained_elapsed                    %            0
sm__mio_pq_write_cycles_active_pipe_lsu.avg.pct_of_peak_sustained_elapsed           %            0
sm__mio_pq_write_cycles_active_pipe_tex.avg.pct_of_peak_sustained_elapsed           %            0
sm__mioc_inst_issued.avg.pct_of_peak_sustained_elapsed                              %         0.39
sm__pipe_fma_cycles_active.avg.pct_of_peak_sustained_elapsed                        %        99.37
sm__pipe_fmaheavy_cycles_active.avg.pct_of_peak_sustained_elapsed                   %        99.37
sm__pipe_fp64_cycles_active.avg.pct_of_peak_sustained_elapsed                       %            0
smsp__inst_issued.sum                                                            inst  115,136,000
------------------------------------------------------------------------- ----------- ------------

*猜测,sm__pipe_fp64_cycles_active并不是那2个FP64 core的metrics,而是smsp里通往fp64 core的接口模块的活动cycle数

*4个subcore里的fp64接口模块,连接到2个fp64 core,并且经过了mio模块.因此,无法打满fp64的利用率

四.进一步证明pipe_fp64_cycles_active并不是2个fp64 core的metrics

c 复制代码
tee fp64_test.cu<<-'EOF'
#include <iostream>
#include <cuda_runtime.h>
#include <iostream>
#include <vector>
#include <stdio.h>
#include <assert.h>
#include <cstdio>
#include <cuda.h>

__global__ void kernel_add_double(volatile double *input,volatile double *output)
{
    unsigned int tid  = threadIdx.x + blockIdx.x * blockDim.x;
    double left=input[tid];
    double right=output[tid];
    for(int i=0;i<1;i++)
    {
        left+=right;
    }       
    output[tid]=left;
}
EOF

/usr/local/cuda/bin/nvcc -std=c++17 -dc -lineinfo -arch=sm_86 -ptx fp64_test.cu -o fp64_test.ptx
/usr/local/cuda/bin/nvcc -arch=sm_86 fp64_test.ptx -cubin -o fp64_test.cubin
/usr/local/cuda/bin/nvcc -arch=sm_86 fp64_test.cubin -fatbin -o fp64_test.fatbin
cat fp64_test.ptx
/usr/local/cuda/bin/cuobjdump --dump-sass fp64_test.fatbin

cuasm.py fp64_test.cubin fp64_test.cuasm
sed '/MOV/d' -i fp64_test.cuasm
sed '/S2R/d' -i fp64_test.cuasm
sed '/ULDC/d' -i fp64_test.cuasm
sed '/IMAD/d' -i fp64_test.cuasm
sed '/LDG/d' -i fp64_test.cuasm
sed '/STG/d' -i fp64_test.cuasm
sed '/F2F/d' -i fp64_test.cuasm

cuasm.py fp64_test.cuasm
/usr/local/cuda/bin/nvcc -arch=sm_86 fp64_test.cubin -fatbin -o fp64_test.fatbin
/usr/local/cuda/bin/cuobjdump --dump-sass fp64_test.fatbin
/usr/local/cuda/bin/cuobjdump --dump-resource-usage fp64_test.fatbin

tee fp64_test_main.cpp<<-'EOF'
#include <stdio.h>
#include <string.h>
#include <cuda_runtime.h>
#include <cuda.h>

int main(int argc,char *argv[])
{
    CUresult error;
    CUdevice cuDevice;
    cuInit(0);
    int deviceCount = 0;
    error = cuDeviceGetCount(&deviceCount);
    error = cuDeviceGet(&cuDevice, 0);
    if(error!=CUDA_SUCCESS)
        {
        printf("Error happened in get device!\n");
    }
    CUcontext cuContext;
    error = cuCtxCreate(&cuContext, 0, cuDevice);
    if(error!=CUDA_SUCCESS)
        {
        printf("Error happened in create context!\n");
    }
    int block_count=1;int block_size=32*4*4;
    int thread_size=block_count*block_size;

    int data_size=sizeof(double)*thread_size;

    double *output_ptr=nullptr;
    double *input_ptr=nullptr;
    int cudaStatus=0;
    cudaStatus = cudaMalloc((void**)&input_ptr, data_size);
    cudaStatus = cudaMalloc((void**)&output_ptr, data_size);
    void *kernelParams[]= {(void*)&output_ptr, (void*)&input_ptr};

    CUmodule module;
    CUfunction double_function;
    const char* module_file = "fp64_test.fatbin";
    const char* double_kernel_name = "_Z17kernel_add_doublePVdS0_";
    
    error = cuModuleLoad(&module, module_file);
    if(error!=CUDA_SUCCESS)
        {
        printf("Error happened in load moudle %d!\n",error);
    }
    error = cuModuleGetFunction(&double_function, module, double_kernel_name);
    if(error!=CUDA_SUCCESS)
    {
        printf("get float_kernel_name error!\n");
    }    
    cuLaunchKernel(double_function,block_count, 1, 1,
                    8, 1, 1,0,0,kernelParams, 0);
    cuLaunchKernel(double_function,block_count, 1, 1,
                    16, 1, 1,0,0,kernelParams, 0);
    cuLaunchKernel(double_function,block_count, 1, 1,
                    32, 1, 1,0,0,kernelParams, 0);
    cuLaunchKernel(double_function,block_count, 1, 1,
                    32*2, 1, 1,0,0,kernelParams, 0);
    cuLaunchKernel(double_function,block_count, 1, 1,
                    32*4, 1, 1,0,0,kernelParams, 0);
    cuLaunchKernel(double_function,block_count, 1, 1,
                    32*5, 1, 1,0,0,kernelParams, 0);
    cuLaunchKernel(double_function,block_count, 1, 1,
                    32*4*8, 1, 1,0,0,kernelParams, 0);
    cudaFree(output_ptr);
    cudaFree(input_ptr);
    cuModuleUnload(module);
    cuCtxDestroy(cuContext);
    return 0;
}
EOF
g++ fp64_test_main.cpp -o fp64_test_main -I /usr/local/cuda/include -L /usr/local/cuda/lib64 -lcudart -lcuda

/usr/local/NVIDIA-Nsight-Compute/ncu --metrics smsp__pipe_fp64_cycles_active ./fp64_test_main

输出

bash 复制代码
kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(8, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         0.14
smsp__pipe_fp64_cycles_active.max       cycle           16
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle           16
--------------------------------- ----------- ------------

kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(16, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         0.14
smsp__pipe_fp64_cycles_active.max       cycle           16  #不足一个warp跟一个warp 的pipe_fp64_cycles_active一样,说明存在无效计算
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle           16
--------------------------------- ----------- ------------

kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(32, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         0.14
smsp__pipe_fp64_cycles_active.max       cycle           16
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle           16
--------------------------------- ----------- ------------

kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(64, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         0.29
smsp__pipe_fp64_cycles_active.max       cycle           16
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle           32
--------------------------------- ----------- ------------

kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(128, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         0.57
smsp__pipe_fp64_cycles_active.max       cycle           16
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle           64
--------------------------------- ----------- ------------

kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(160, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         0.71
smsp__pipe_fp64_cycles_active.max       cycle           32
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle           80
--------------------------------- ----------- ------------

kernel_add_double(volatile double *, volatile double *) (1, 1, 1)x(1024, 1, 1), Context 1, Stream 7, Device 0, CC 8.6
Section: Command line profiler metrics
--------------------------------- ----------- ------------
Metric Name                       Metric Unit Metric Value
--------------------------------- ----------- ------------
smsp__pipe_fp64_cycles_active.avg       cycle         4.57
smsp__pipe_fp64_cycles_active.max       cycle          128
smsp__pipe_fp64_cycles_active.min       cycle            0
smsp__pipe_fp64_cycles_active.sum       cycle          512 #每个smsp 执行一个warp的fp64需要16个pipe_fp64_cycles_active
--------------------------------- ----------- ------------
  • 如果是2个fp64 cores的metrics,不会出现这样的现象
相关推荐
前端青山7 小时前
Node.js-增强 API 安全性和性能优化
开发语言·前端·javascript·性能优化·前端框架·node.js
青云交13 小时前
大数据新视界 -- 大数据大厂之 Impala 性能优化:应对海量复杂数据的挑战(上)(7/30)
大数据·性能优化·impala·数据分区·查询优化·海量复杂数据·经典案例
chusheng184016 小时前
Python 爬取大量数据如何并发抓取与性能优化
开发语言·python·性能优化
探索云原生1 天前
GPU 环境搭建指南:如何在裸机、Docker、K8s 等环境中使用 GPU
ai·云原生·kubernetes·go·gpu
XMYX-01 天前
MySQL 性能优化策略:提升响应速度与系统稳定性
mysql·性能优化
一个处女座的程序猿1 天前
AI之硬件对比:据传英伟达Nvidia2025年将推出RTX 5090-32GB/RTX 5080-24GB、华为2025年推出910C/910D
人工智能·gpu
PangPiLoLo2 天前
高可用架构-业务高可用
java·性能优化·架构
尸僵打怪兽2 天前
软考(中级-软件设计师)数据库篇(1101)
数据库·oracle·性能优化·软考
程序猿进阶2 天前
系统上云-流量分析和链路分析
java·后端·阿里云·面试·性能优化·系统架构·云计算
飞腾开发者3 天前
飞腾平台Arm ComputeLibrary编译安装指南
linux·服务器·arm开发·后端·性能优化