Row-wise/column-wise operations on matrices with CUDA - c++

Question

I'm relatively new to CUDA programming. I have understood the programming model and have already written few basic kernels. I know how to apply a kernel to each element of a matrix (stored as 1D array), but now I'm trying to figure out how to apply the same operation to the same row/column of the input matrix.
Let's say I have a MxN matrix and a vector of length N. I would like to sum (but it can be any other math operation) the vector to each row of the matrix.
The serial code of such operation is:
for (int c = 0; c < columns; c++)
{
for (int r = 0; r < rows; r++)
{
M[r * rows + c] += V[c];
}
}
Now the CUDA code for doing this operation should be quite straightforward: I should spawn as many cuda threads as the elements and apply this kernel:
__global__ void kernel(const unsigned int size, float* matrix, const float* vector)
{
// get the current element index for the thread
unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < size)
{
// sum the current element with the
matrix[idx] += vector[threadIdx.x];
}
}
It runs but the result is not correct. Actually, it's correct if I transpose the matrix after the kernel completes its work. Unfortunately, I have no clue why it works in this way. Could you help me to figure out this problem? Thanks in advance.
EDIT #1
I launch the kernel using:
int block_size = 64;
int grid_size = (M * N + block_size - 1) / block_size;
kernel<<<grid_size, block_size>>>(M * N, matrix, vector);
EDIT #2
I solved the problem by fixing the CPU code as suggested by #RobertCrovella:
M[r * columns + c] += V[c];
It should match the outer for, that is, over the columns.

Answer 1

The kernel shown in the question could be used without modification to sum a vector to each of the rows of a matrix (assuming c-style row-major storage), subject to certain limitations. A demonstration is here.
The main limitation of that approach is that the maximum vector length and therefore matrix width that can be handled is equal to the maximum number of threads per block, which on current CUDA 7-supported GPUs is 1024.
We can eliminate that limitation with a slight modification to the vector indexing, and passing the row width (number of columns) as a parameter to the matrix. With this modification, we should be able to handle arbitrary matrix (and vector) sizes.
EDIT: based on discussion/comments, OP wants to know how to handle row-major or column major underlying storage. The following example uses a templated kernel to select either row-major or column major underlying storage, and also shows one possible CUBLAS method for doing a add-vector-to-each-matrix-row operation using rank-1 update function:
$ cat t712.cu
#include <iostream>
#include <cublas_v2.h>
#define ROWS 20
#define COLS 10
#define nTPB 64
#define ROW_MAJOR 0
#define COL_MAJOR 1
template <int select, typename T>
__global__ void vec_mat_row_add(const unsigned int height, const unsigned int width, T* matrix, const T* vector)
{
// get the current element index for the thread
unsigned int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < height*width)
{
// sum the current element with the
if (select == ROW_MAJOR)
matrix[idx] += vector[idx%width];
else // COL_MAJOR
matrix[idx] += vector[idx/height];
}
}
int main(){
float *h_mat, *d_mat, *h_vec, *d_vec;
const unsigned int msz = ROWS*COLS*sizeof(float);
const unsigned int vsz = COLS*sizeof(float);
h_mat = (float *)malloc(msz);
h_vec = (float *)malloc(vsz);
cudaMalloc(&d_mat, msz);
cudaMalloc(&d_vec, vsz);
for (int i=0; i<COLS; i++) h_vec[i] = i; // set vector to 0,1,2, ...
cudaMemcpy(d_vec, h_vec, vsz, cudaMemcpyHostToDevice);
// test row-major case
cudaMemset(d_mat, 0, msz); // set matrix to zero
vec_mat_row_add<ROW_MAJOR><<<(ROWS*COLS + nTPB -1)/nTPB, nTPB>>>(ROWS, COLS, d_mat, d_vec);
cudaMemcpy(h_mat, d_mat, msz, cudaMemcpyDeviceToHost);
std::cout << "Row-major result: " << std::endl;
for (int i = 0; i < ROWS; i++){
for (int j = 0; j < COLS; j++) std::cout << h_mat[i*COLS+j] << " ";
std::cout << std::endl;}
// test column-major case
cudaMemset(d_mat, 0, msz); // set matrix to zero
vec_mat_row_add<COL_MAJOR><<<(ROWS*COLS + nTPB -1)/nTPB, nTPB>>>(ROWS, COLS, d_mat, d_vec);
cudaMemcpy(h_mat, d_mat, msz, cudaMemcpyDeviceToHost);
std::cout << "Column-major result: " << std::endl;
for (int i = 0; i < ROWS; i++){
for (int j = 0; j < COLS; j++) std::cout << h_mat[j*ROWS+i] << " ";
std::cout << std::endl;}
// test CUBLAS, doing matrix-vector add using <T>ger
cudaMemset(d_mat, 0, msz); // set matrix to zero
float *d_ones, *h_ones;
h_ones = (float *)malloc(ROWS*sizeof(float));
for (int i =0; i<ROWS; i++) h_ones[i] = 1.0f;
cudaMalloc(&d_ones, ROWS*sizeof(float));
cudaMemcpy(d_ones, h_ones, ROWS*sizeof(float), cudaMemcpyHostToDevice);
cublasHandle_t ch;
cublasCreate(&ch);
float alpha = 1.0f;
cublasStatus_t stat = cublasSger(ch, ROWS, COLS, &alpha, d_ones, 1, d_vec, 1, d_mat, ROWS);
if (stat != CUBLAS_STATUS_SUCCESS) {std::cout << "CUBLAS error: " << (int)stat << std::endl; return 1;}
cudaMemcpy(h_mat, d_mat, msz, cudaMemcpyDeviceToHost);
std::cout << "CUBLAS Column-major result: " << std::endl;
for (int i = 0; i < ROWS; i++){
for (int j = 0; j < COLS; j++) std::cout << h_mat[j*ROWS+i] << " ";
std::cout << std::endl;}
return 0;
}
$ nvcc -o t712 t712.cu -lcublas
$ ./t712
Row-major result:
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
Column-major result:
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
CUBLAS Column-major result:
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
0 1 2 3 4 5 6 7 8 9
$
For brevity of presentation, I've not included proper cuda error checking, but that is always a good idea any time you are having trouble with a CUDA code. As a proxy/shortcut, you can run your code with cuda-memcheck as a quick check to see if there are any CUDA errors.
Note that we expect all 3 printouts to be identical because that is actually the correct way to display the matrix, regardless of whether the underlying storage is row-major or column-major. The difference in underlying storage is accounted for in the for-loops handling the display output.

Answer 2

Robert Crovella has already answered this question providing examples using explicit CUDA kernels and cuBLAS.
I find it useful, for future references, to show also an example on how performing row-wise or column-wise operations using CUDA Thrust. In particular, I'm focusing on two problems:
Summing a column vector to all matrix columns;
Summing a row vector to all matrix rows.
The generality of thrust::transform enables to generalize the example below to elementwise operations other than the sum (e.g., multiplications, divisions, subtractions etc.).
#include <thrust/device_vector.h>
#include <thrust/reduce.h>
#include <thrust/random.h>
#include <thrust/sort.h>
#include <thrust/unique.h>
#include <thrust/equal.h>
using namespace thrust::placeholders;
/*************************************/
/* CONVERT LINEAR INDEX TO ROW INDEX */
/*************************************/
template <typename T>
struct linear_index_to_row_index : public thrust::unary_function<T,T> {
T Ncols; // --- Number of columns
__host__ __device__ linear_index_to_row_index(T Ncols) : Ncols(Ncols) {}
__host__ __device__ T operator()(T i) { return i / Ncols; }
};
/********/
/* MAIN */
/********/
int main()
{
/**************************/
/* SETTING UP THE PROBLEM */
/**************************/
const int Nrows = 10; // --- Number of rows
const int Ncols = 3; // --- Number of columns
// --- Random uniform integer distribution between 0 and 100
thrust::default_random_engine rng;
thrust::uniform_int_distribution<int> dist1(0, 100);
// --- Random uniform integer distribution between 1 and 4
thrust::uniform_int_distribution<int> dist2(1, 4);
// --- Matrix allocation and initialization
thrust::device_vector<float> d_matrix(Nrows * Ncols);
for (size_t i = 0; i < d_matrix.size(); i++) d_matrix[i] = (float)dist1(rng);
// --- Column vector allocation and initialization
thrust::device_vector<float> d_column(Nrows);
for (size_t i = 0; i < d_column.size(); i++) d_column[i] = (float)dist2(rng);
// --- Row vector allocation and initialization
thrust::device_vector<float> d_row(Ncols);
for (size_t i = 0; i < d_row.size(); i++) d_row[i] = (float)dist2(rng);
printf("\n\nOriginal matrix\n");
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix[i * Ncols + j] << " ";
std::cout << "]\n";
}
printf("\n\nColumn vector\n");
for(int i = 0; i < Nrows; i++) std::cout << d_column[i] << "\n";
printf("\n\nRow vector\n");
for(int i = 0; i < Ncols; i++) std::cout << d_row[i] << " ";
/*******************************************************/
/* ADDING THE SAME COLUMN VECTOR TO ALL MATRIX COLUMNS */
/*******************************************************/
thrust::device_vector<float> d_matrix2(d_matrix);
thrust::transform(d_matrix.begin(), d_matrix.end(),
thrust::make_permutation_iterator(
d_column.begin(),
thrust::make_transform_iterator(thrust::make_counting_iterator(0), linear_index_to_row_index<int>(Ncols))),
d_matrix2.begin(),
thrust::plus<float>());
printf("\n\nColumn + Matrix -> Result matrix\n");
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix2[i * Ncols + j] << " ";
std::cout << "]\n";
}
/*************************************************/
/* ADDING THE SAME ROW VECTOR TO ALL MATRIX ROWS */
/*************************************************/
thrust::device_vector<float> d_matrix3(d_matrix);
thrust::transform(thrust::make_permutation_iterator(
d_matrix.begin(),
thrust::make_transform_iterator(thrust::make_counting_iterator(0),(_1 % Nrows) * Ncols + _1 / Nrows)),
thrust::make_permutation_iterator(
d_matrix.begin(),
thrust::make_transform_iterator(thrust::make_counting_iterator(0),(_1 % Nrows) * Ncols + _1 / Nrows)) + Nrows * Ncols,
thrust::make_permutation_iterator(
d_row.begin(),
thrust::make_transform_iterator(thrust::make_counting_iterator(0), linear_index_to_row_index<int>(Nrows))),
thrust::make_permutation_iterator(
d_matrix3.begin(),
thrust::make_transform_iterator(thrust::make_counting_iterator(0),(_1 % Nrows) * Ncols + _1 / Nrows)),
thrust::plus<float>());
printf("\n\nRow + Matrix -> Result matrix\n");
for(int i = 0; i < Nrows; i++) {
std::cout << "[ ";
for(int j = 0; j < Ncols; j++)
std::cout << d_matrix3[i * Ncols + j] << " ";
std::cout << "]\n";
}
return 0;
}