CUBLAS 矩阵乘法与行主数据无转置
CUBLAS matrix multiplication with row-major data without transpose
我目前正在尝试在我的 GPU 上使用 CUBLAS 实现矩阵乘法。
它适用于方阵和特定大小的输入,但对于其他输入,最后一行不返回(并且包含 0,因为这是我实现它的方式) .
我认为这是 cublasSgemm
的分配或语法问题,但我找不到它在哪里。
N.B. :如果您不熟悉 CUBLAS:
它是 column-majored,这就是为什么看起来操作是以另一种方式执行的。
如有任何帮助,我们将不胜感激。
代码:
请注意,gpuErrchk
和cublasErrchk
当然与这里无关。
#include <cuda.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <vector>
std::vector<float> CUDA_mult_MAT(const std::vector<float> &data_1 , const uint64_t data_1_rows, const uint64_t data_1_columns,
const std::vector<float> &data_2 , const uint64_t data_2_rows, const uint64_t data_2_columns){
cublasHandle_t handle;
cublasErrchk(cublasCreate(&handle));
std::vector<float> result(data_1_rows * data_2_columns); //Vector holding the result of the multiplication
/*----------------------------------------------------------------------------------------------*/
float* GPU_data_1 = NULL;
gpuErrchk(cudaMalloc((void**)&GPU_data_1 , data_1.size()*sizeof(float))); //Allocate memory on the GPU
gpuErrchk(cudaMemcpy(GPU_data_1, data_1.data(), data_1.size()*sizeof(float), cudaMemcpyHostToDevice)); //Copy data from data_1 to GPU_data_1
float* GPU_data_2 = NULL;
gpuErrchk(cudaMalloc((void**)&GPU_data_2 ,data_2.size()*sizeof(float))); //Allocate memory on the GPU
gpuErrchk(cudaMemcpy(GPU_data_2, data_2.data(), data_2.size()*sizeof(float), cudaMemcpyHostToDevice));//Copy data from data_2 to GPU_data_2
float* GPU_result = NULL;
gpuErrchk(cudaMalloc((void**)&GPU_result , result.size()*sizeof(float))); //Allocate memory on the GPU
/*----------------------------------------------------------------------------------------------*/
const float alpha = 1.f;
const float beta = 0.f;
cublasErrchk(
cublasSgemm(handle , CUBLAS_OP_N , CUBLAS_OP_N,
data_2_columns , data_2_rows ,data_1_columns,
&alpha , GPU_data_2 , data_2_columns,
GPU_data_1 , data_1_columns,
&beta , GPU_result , data_1_rows)
); //Perform multiplication
gpuErrchk(cudaMemcpy(result.data() , GPU_result , result.size() * sizeof(float) , cudaMemcpyDeviceToHost)); //Copy back to the vector 'result'
gpuErrchk(cudaFree(GPU_data_1)); //Free GPU memory
gpuErrchk(cudaFree(GPU_data_2)); //Free GPU memory
gpuErrchk(cudaFree(GPU_result)); //Free GPU memory
cublasErrchk(cublasDestroy_v2(handle));
return result;
}
输入:
#include <iostream>
#include <vector>
int main(){
const std::vector<float> r1 = CUDA_mult_MAT({1 , 2 , 3 , 4 , 5 , 6} , 2 , 3 ,
{7 , 8 , 9 , 10 , 11 , 12} , 3 , 2);
/*
Product :
7 8
1 2 3 9 10
4 5 6 11 12
*/
for(auto & value: r1){std::cout << value << " " ;}
std::cout << std::endl;
const std::vector<float> r2 = CUDA_mult_MAT({7 , 8 , 9 , 10 , 11 , 12} , 3 , 2 ,
{1 , 2 , 3 , 4 , 5 , 6} , 2 , 3);
/*
Product :
7 8
9 10 1 2 3
11 12 4 5 6
*/
for(auto & value: r2){std::cout << value << " " ;}
std::cout << std::endl;
return 0;
}
输出:
程序打印:
58 64 139 154
39 54 69 49 68 87 0 0 0
^~~~~~~
预计:
58 64 139 154
39 54 69 49 68 87 59 82 105
^~~~~~~
我们可以通过不同方式观察到您使用 CUBLAS 的问题。
首先研究CUBLAS Sgemm documentation,我们看到3个参数m
,n
,k
出现,依次 紧跟在转置说明符之后:
cublasStatus_t cublasSgemm(cublasHandle_t handle,
cublasOperation_t transa, cublasOperation_t transb,
int m, int n, int k,
^ ^ ^
我们还观察到矩阵维度由下式给出:
A , B and C are matrices stored in column-major format with dimensions op ( A ) m × k , op ( B ) k × n and C m × n ,
所以第一个输入矩阵的维度是m x k
第二个输入矩阵的维度为k x n
,输出矩阵的维度为m x n
让我们暂时关注输出矩阵。鉴于其尺寸是使用 m
和 n
参数指定的,它不可能是正确的(假设在非正方形的情况下)仅传递 data_2
尺寸:
cublasSgemm(handle , CUBLAS_OP_N , CUBLAS_OP_N,
data_2_columns , data_2_rows ,data_1_columns,
^^^^^^^^^^^^^^ ^^^^^^^^^^^
其次,从错误检查的角度来看,您可以通过 运行 使用 cuda-memcheck
来快速估计您的 CUBLAS 调用有问题。报告的第一个错误如下:
$ cuda-memcheck ./t23
========= CUDA-MEMCHECK
========= Invalid __global__ read of size 4
========= at 0x000006f0 in void gemmSN_NN_kernel<float, int=256, int=4, int=2, int=8, int=3, int=4, bool=0, cublasGemvTensorStridedBatched<float const >, cublasGemvTensorStridedBatched<float>>(cublasGemmSmallNParams<float const , cublasGemvTensorStridedBatched<float const >, float>)
========= by thread (64,0,0) in block (0,0,0)
========= Address 0x7f9c30a2061c is out of bounds
========= Device Frame:void gemmSN_NN_kernel<float, int=256, int=4, int=2, int=8, int=3, int=4, bool=0, cublasGemvTensorStridedBatched<float const >, cublasGemvTensorStridedBatched<float>>(cublasGemmSmallNParams<float const , cublasGemvTensorStridedBatched<float const >, float>) (void gemmSN_NN_kernel<float, int=256, int=4, int=2, int=8, int=3, int=4, bool=0, cublasGemvTensorStridedBatched<float const >, cublasGemvTensorStridedBatched<float>>(cublasGemmSmallNParams<float const , cublasGemvTensorStridedBatched<float const >, float>) : 0x6f0)
========= Saved host backtrace up to driver entry point at kernel launch time
========= Host Frame:/usr/lib/x86_64-linux-gnu/libcuda.so.1 (cuLaunchKernel + 0x2b8) [0x1e5cc8]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x1063c8b]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x10a9965]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x6bfacc]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x5fc7af]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x436c35]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 (cublasLtMatmul + 0x60f) [0x43484f]
========= Host Frame:/usr/local/cuda/lib64/libcublas.so.11 [0x9ef6db]
========= Host Frame:/usr/local/cuda/lib64/libcublas.so.11 [0x50e4f0]
========= Host Frame:/usr/local/cuda/lib64/libcublas.so.11 (cublasSgemm_v2 + 0x1ee) [0x50f29e]
========= Host Frame:./t23 [0x7986]
========= Host Frame:./t23 [0x7b4c]
========= Host Frame:/lib/x86_64-linux-gnu/libc.so.6 (__libc_start_main + 0xe7) [0x21b97]
========= Host Frame:./t23 [0x744a]
=========
当然,一种可能的解决方案是转置输入矩阵,因此它们按列主要顺序排列,CUBLAS 提供了 Sgemm
选项来做到这一点(见上文)。但是在我看来,您要做的是在不转置输入数组的情况下进行 C 风格的行主乘法。有一篇文章 here 描述了如何做到这一点。
当我将该启发式应用于您的 cublasSgemm()
调用时,我得到了:
cublasSgemm(handle , CUBLAS_OP_N , CUBLAS_OP_N,
data_2_columns , data_1_rows ,data_1_columns,
&alpha , GPU_data_2 , data_2_columns,
GPU_data_1 , data_1_columns,
&beta , GPU_result , data_2_columns)
当我编译并运行你的代码进行这些更改时,我得到这个:
$ cuda-memcheck ./t23
========= CUDA-MEMCHECK
58 64 139 154
39 54 69 49 68 87 59 82 105
========= ERROR SUMMARY: 0 errors
我目前正在尝试在我的 GPU 上使用 CUBLAS 实现矩阵乘法。
它适用于方阵和特定大小的输入,但对于其他输入,最后一行不返回(并且包含 0,因为这是我实现它的方式) .
我认为这是 cublasSgemm
的分配或语法问题,但我找不到它在哪里。
N.B. :如果您不熟悉 CUBLAS: 它是 column-majored,这就是为什么看起来操作是以另一种方式执行的。
如有任何帮助,我们将不胜感激。
代码:
请注意,gpuErrchk
和cublasErrchk
当然与这里无关。
#include <cuda.h>
#include <cuda_runtime.h>
#include <cublas_v2.h>
#include <vector>
std::vector<float> CUDA_mult_MAT(const std::vector<float> &data_1 , const uint64_t data_1_rows, const uint64_t data_1_columns,
const std::vector<float> &data_2 , const uint64_t data_2_rows, const uint64_t data_2_columns){
cublasHandle_t handle;
cublasErrchk(cublasCreate(&handle));
std::vector<float> result(data_1_rows * data_2_columns); //Vector holding the result of the multiplication
/*----------------------------------------------------------------------------------------------*/
float* GPU_data_1 = NULL;
gpuErrchk(cudaMalloc((void**)&GPU_data_1 , data_1.size()*sizeof(float))); //Allocate memory on the GPU
gpuErrchk(cudaMemcpy(GPU_data_1, data_1.data(), data_1.size()*sizeof(float), cudaMemcpyHostToDevice)); //Copy data from data_1 to GPU_data_1
float* GPU_data_2 = NULL;
gpuErrchk(cudaMalloc((void**)&GPU_data_2 ,data_2.size()*sizeof(float))); //Allocate memory on the GPU
gpuErrchk(cudaMemcpy(GPU_data_2, data_2.data(), data_2.size()*sizeof(float), cudaMemcpyHostToDevice));//Copy data from data_2 to GPU_data_2
float* GPU_result = NULL;
gpuErrchk(cudaMalloc((void**)&GPU_result , result.size()*sizeof(float))); //Allocate memory on the GPU
/*----------------------------------------------------------------------------------------------*/
const float alpha = 1.f;
const float beta = 0.f;
cublasErrchk(
cublasSgemm(handle , CUBLAS_OP_N , CUBLAS_OP_N,
data_2_columns , data_2_rows ,data_1_columns,
&alpha , GPU_data_2 , data_2_columns,
GPU_data_1 , data_1_columns,
&beta , GPU_result , data_1_rows)
); //Perform multiplication
gpuErrchk(cudaMemcpy(result.data() , GPU_result , result.size() * sizeof(float) , cudaMemcpyDeviceToHost)); //Copy back to the vector 'result'
gpuErrchk(cudaFree(GPU_data_1)); //Free GPU memory
gpuErrchk(cudaFree(GPU_data_2)); //Free GPU memory
gpuErrchk(cudaFree(GPU_result)); //Free GPU memory
cublasErrchk(cublasDestroy_v2(handle));
return result;
}
输入:
#include <iostream>
#include <vector>
int main(){
const std::vector<float> r1 = CUDA_mult_MAT({1 , 2 , 3 , 4 , 5 , 6} , 2 , 3 ,
{7 , 8 , 9 , 10 , 11 , 12} , 3 , 2);
/*
Product :
7 8
1 2 3 9 10
4 5 6 11 12
*/
for(auto & value: r1){std::cout << value << " " ;}
std::cout << std::endl;
const std::vector<float> r2 = CUDA_mult_MAT({7 , 8 , 9 , 10 , 11 , 12} , 3 , 2 ,
{1 , 2 , 3 , 4 , 5 , 6} , 2 , 3);
/*
Product :
7 8
9 10 1 2 3
11 12 4 5 6
*/
for(auto & value: r2){std::cout << value << " " ;}
std::cout << std::endl;
return 0;
}
输出:
程序打印:
58 64 139 154
39 54 69 49 68 87 0 0 0
^~~~~~~
预计:
58 64 139 154
39 54 69 49 68 87 59 82 105
^~~~~~~
我们可以通过不同方式观察到您使用 CUBLAS 的问题。
首先研究CUBLAS Sgemm documentation,我们看到3个参数m
,n
,k
出现,依次 紧跟在转置说明符之后:
cublasStatus_t cublasSgemm(cublasHandle_t handle,
cublasOperation_t transa, cublasOperation_t transb,
int m, int n, int k,
^ ^ ^
我们还观察到矩阵维度由下式给出:
A , B and C are matrices stored in column-major format with dimensions op ( A ) m × k , op ( B ) k × n and C m × n ,
所以第一个输入矩阵的维度是m x k
第二个输入矩阵的维度为k x n
,输出矩阵的维度为m x n
让我们暂时关注输出矩阵。鉴于其尺寸是使用 m
和 n
参数指定的,它不可能是正确的(假设在非正方形的情况下)仅传递 data_2
尺寸:
cublasSgemm(handle , CUBLAS_OP_N , CUBLAS_OP_N,
data_2_columns , data_2_rows ,data_1_columns,
^^^^^^^^^^^^^^ ^^^^^^^^^^^
其次,从错误检查的角度来看,您可以通过 运行 使用 cuda-memcheck
来快速估计您的 CUBLAS 调用有问题。报告的第一个错误如下:
$ cuda-memcheck ./t23
========= CUDA-MEMCHECK
========= Invalid __global__ read of size 4
========= at 0x000006f0 in void gemmSN_NN_kernel<float, int=256, int=4, int=2, int=8, int=3, int=4, bool=0, cublasGemvTensorStridedBatched<float const >, cublasGemvTensorStridedBatched<float>>(cublasGemmSmallNParams<float const , cublasGemvTensorStridedBatched<float const >, float>)
========= by thread (64,0,0) in block (0,0,0)
========= Address 0x7f9c30a2061c is out of bounds
========= Device Frame:void gemmSN_NN_kernel<float, int=256, int=4, int=2, int=8, int=3, int=4, bool=0, cublasGemvTensorStridedBatched<float const >, cublasGemvTensorStridedBatched<float>>(cublasGemmSmallNParams<float const , cublasGemvTensorStridedBatched<float const >, float>) (void gemmSN_NN_kernel<float, int=256, int=4, int=2, int=8, int=3, int=4, bool=0, cublasGemvTensorStridedBatched<float const >, cublasGemvTensorStridedBatched<float>>(cublasGemmSmallNParams<float const , cublasGemvTensorStridedBatched<float const >, float>) : 0x6f0)
========= Saved host backtrace up to driver entry point at kernel launch time
========= Host Frame:/usr/lib/x86_64-linux-gnu/libcuda.so.1 (cuLaunchKernel + 0x2b8) [0x1e5cc8]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x1063c8b]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x10a9965]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x6bfacc]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x5fc7af]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 [0x436c35]
========= Host Frame:/usr/local/cuda/lib64/libcublasLt.so.11 (cublasLtMatmul + 0x60f) [0x43484f]
========= Host Frame:/usr/local/cuda/lib64/libcublas.so.11 [0x9ef6db]
========= Host Frame:/usr/local/cuda/lib64/libcublas.so.11 [0x50e4f0]
========= Host Frame:/usr/local/cuda/lib64/libcublas.so.11 (cublasSgemm_v2 + 0x1ee) [0x50f29e]
========= Host Frame:./t23 [0x7986]
========= Host Frame:./t23 [0x7b4c]
========= Host Frame:/lib/x86_64-linux-gnu/libc.so.6 (__libc_start_main + 0xe7) [0x21b97]
========= Host Frame:./t23 [0x744a]
=========
当然,一种可能的解决方案是转置输入矩阵,因此它们按列主要顺序排列,CUBLAS 提供了 Sgemm
选项来做到这一点(见上文)。但是在我看来,您要做的是在不转置输入数组的情况下进行 C 风格的行主乘法。有一篇文章 here 描述了如何做到这一点。
当我将该启发式应用于您的 cublasSgemm()
调用时,我得到了:
cublasSgemm(handle , CUBLAS_OP_N , CUBLAS_OP_N,
data_2_columns , data_1_rows ,data_1_columns,
&alpha , GPU_data_2 , data_2_columns,
GPU_data_1 , data_1_columns,
&beta , GPU_result , data_2_columns)
当我编译并运行你的代码进行这些更改时,我得到这个:
$ cuda-memcheck ./t23
========= CUDA-MEMCHECK
58 64 139 154
39 54 69 49 68 87 59 82 105
========= ERROR SUMMARY: 0 errors