[英]CUDA: Filling a column-major matrix
我對CUDA相當陌生,並且我正嘗試將我為性能關鍵型項目所做的一些繁瑣的計算工作卸載到GPU。 在我的計算機上,我有兩張NVS 510圖形卡,但是我目前只在嘗試其中一張。
我有一些要填充的大型列矩陣(1000-5000行x 1-5 M列)。 到目前為止,我能夠像編寫數組一樣編寫代碼來填充矩陣,並且它對於較小尺寸的矩陣也能很好地工作。
__global__ void interp_kernel(fl_type * d_matrix, fl_type* weights, [other params],
int n_rows, int num_cols) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int column = index / n_rows;
int row = index % n_rows;
if (row > n_sim || column > num_cols) return;
d_matrix[index] = …something(row, column,[other params]);
}
內核稱為:
fl_type *res;
cudaMalloc((void**)&res, n_columns*n_rows*fl_size);
int block_size = 1024;
int num_blocks = (n_rows* n_columns + block_size - 1) / block_size;
std::cout << "num_blocks:" << num_blocks << std::endl;
interp_kernel << < num_blocks, block_size >> > (res,[other params], n_rows,n_columns);
一切都很好。 如果我更改內核以使用2D線程:
__global__ void interp_kernel2D(fl_type * d_matrix, fl_type* weights, [other params],
int n_rows, int num_cols) {
int column = blockIdx.x * blockDim.x + threadIdx.x;
int row = blockIdx.y * blockDim.y + threadIdx.y;
int index = column* n_rows + row;
if (row > n_rows || column > num_cols) return;
d_matrix[index] = …something(row, column,[other params]);
}
我調用它
int block_size2 = 32; //each block will have block_size2*block_size2 threads
dim3 num_blocks2(block_size2, block_size2);
int x_grid = (n_columns + block_size2 - 1) / block_size2;
int y_grid = (n_rows + block_size2 - 1) / block_size2;
dim3 grid_size2(x_grid, y_grid);
interp_kernel2D <<< grid_size2, num_blocks2 >>> (res,[other params], n_rows,n_columns);
結果全為零,CUDA返回未知錯誤。 我想念什么? 可以在此處找到使用VS2015和CUDA 8.0編譯時沒有錯誤的實際代碼: https ://pastebin.com/XBCVC7VV
這是pastebin鏈接中的代碼:
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <assert.h>
#include <iostream>
#include <random>
#include <chrono>
typedef float fl_type;
typedef int pos_type;
typedef std::chrono::milliseconds ms;
//declaration of the cuda function
void cuda_interpolation_function(fl_type* interp_value_back, int result_size, fl_type * grid_values, int grid_values_size, fl_type* weights, pos_type* node_map, int total_action_number, int interp_dim, int n_sim);
fl_type iterp_cpu(fl_type* weights, pos_type* node_map, fl_type* grid_values, int& row, int& column, int& interp_dim, int& n_sim) {
int w_p = column*interp_dim;
fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
res += weights[w_p + inter_point] * grid_values[node_map[w_p + inter_point] * n_sim + row];
}
return res;
}
__global__ void interp_kernel(fl_type * d_matrix, fl_type* weights, pos_type* node_map, fl_type* grid_values, int interp_dim, int n_sim, int num_cols) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int column = index / n_sim;
int row = index % n_sim;
int w_p = column*interp_dim;
if (row > n_sim || column > num_cols) return;
fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
res += weights[w_p + inter_point] * grid_values[row + node_map[w_p + inter_point] * n_sim];
}
d_matrix[index] = res;
}
__global__ void interp_kernel2D(fl_type * d_matrix, fl_type* weights, pos_type* node_map, fl_type* grid_values, int interp_dim, int n_sim, int num_cols) {
int column = blockIdx.x * blockDim.x + threadIdx.x;
int row = blockIdx.y * blockDim.y + threadIdx.y;
int index = column*n_sim + row;
int w_p = column*interp_dim;
if (row > n_sim || column > num_cols) return;
fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
res += weights[w_p + inter_point] * grid_values[row + node_map[w_p + inter_point] * n_sim];
}
d_matrix[index] = res;
}
void verify(fl_type *host, fl_type *device, int size) {
int count = 0;
int count_zero = 0;
for (int i = 0; i < size; i++) {
if (host[i] != device[i]) {
count++;
//std::cout <<"pos: " <<i<< " CPU:" <<h[i] << ", GPU: " << d[i] <<std::endl;
assert(host[i] == device[i]);
if (device[i] == 0.0)
count_zero++;
}
}
if (count) {
std::cout << "Non matching: " << count << "out of " << size << "(" << (float(count) / size * 100) << "%)" << std::endl;
std::cout << "Zeros returned from the device: " << count_zero <<"(" << (float(count_zero) / size * 100) << "%)" << std::endl;
}
else
std::cout << "Perfect match!" << std::endl;
}
int main() {
int fl_size = sizeof(fl_type);
int pos_size = sizeof(pos_type);
int dim = 5; // range: 2-5
int number_nodes = 5500; // range: 10.000-500.000
int max_actions = 12; // range: 6-200
int n_sim = 1000; // range: 1.000-10.000
int interp_dim = std::pow(2, dim);
int grid_values_size = n_sim*number_nodes;
std::default_random_engine generator;
std::normal_distribution<fl_type> normal_dist(0.0, 1);
std::uniform_int_distribution<> uniform_dist(0, number_nodes - 1);
double bit_allocated = 0;
fl_type * grid_values; //flattened 2d array, containing the value of the grid (n_sims x number_nodes)
grid_values = (fl_type *)malloc(grid_values_size * fl_size);
bit_allocated += grid_values_size * fl_size;
for (int i = 0; i < grid_values_size; i++)
grid_values[i] = normal_dist(generator);
pos_type * map_node2values_start; //vector that maps each node to the first column of the result matrix regarding that done
pos_type * map_node2values_how_many; //vector that stores how many action we have per node
map_node2values_start = (pos_type *)malloc(number_nodes * pos_size);
map_node2values_how_many = (pos_type *)malloc(number_nodes * pos_size);
bit_allocated += 2 * (number_nodes * pos_size);
for (int i = 0; i < number_nodes; i++) {
//each node as simply max_actions
map_node2values_start[i] = max_actions*i;
map_node2values_how_many[i] = max_actions;
}
//total number of actions, which is amount of column of the results
int total_action_number = map_node2values_start[number_nodes - 1] + map_node2values_how_many[number_nodes - 1];
//vector that keep tracks of the columnt to grab, and their weight in the interpolation
fl_type* weights;
pos_type * node_map;
weights = (fl_type *)malloc(total_action_number*interp_dim * pos_size);
bit_allocated += total_action_number * fl_size;
node_map = (pos_type *)malloc(total_action_number*interp_dim * pos_size);
bit_allocated += total_action_number * pos_size;
//filling with random numbers
for (int i = 0; i < total_action_number*interp_dim; i++) {
node_map[i] = uniform_dist(generator); // picking random column
weights[i] = 1.0 / interp_dim; // uniform weights
}
std::cout << "done filling!" << std::endl;
std::cout << bit_allocated / 8 / 1024 / 1024 << "MB allocated" << std::endl;
int result_size = n_sim*total_action_number;
fl_type *interp_value_cpu;
bit_allocated += result_size* fl_size;
interp_value_cpu = (fl_type *)malloc(result_size* fl_size);
auto start = std::chrono::steady_clock::now();
for (int row = 0; row < n_sim; row++) {
for (int column = 0; column < total_action_number; column++) {
auto zz = iterp_cpu(weights, node_map, grid_values, row, column, interp_dim, n_sim);
interp_value_cpu[column*n_sim + row] = zz;
}
}
auto elapsed_cpu = std::chrono::steady_clock::now() - start;
std::cout << "Crunching values on the CPU (serial): " << std::chrono::duration_cast<ms>(elapsed_cpu).count() / 1000.0 << "s" << std::endl;
int * pp;
cudaMalloc((void**)&pp, sizeof(int)); //initializing the device, to not affect the benchmark
fl_type *interp_value_gpu;
interp_value_gpu = (fl_type *)malloc(result_size* fl_size);
start = std::chrono::steady_clock::now();
cuda_interpolation_function(interp_value_gpu, result_size, grid_values, grid_values_size, weights, node_map, total_action_number, interp_dim, n_sim);
auto elapsed_gpu = std::chrono::steady_clock::now() - start;
std::cout << "Crunching values on the GPU: " << std::chrono::duration_cast<ms>(elapsed_gpu).count() / 1000.0 << "s" << std::endl;
float ms_cpu = std::chrono::duration_cast<ms>(elapsed_cpu).count();
float ms_gpu = std::chrono::duration_cast<ms>(elapsed_gpu).count();
int n_proc = 4;
std::cout << "Performance: " << (ms_gpu- ms_cpu / n_proc) / (ms_cpu / n_proc) * 100 << " % less time than parallel CPU!" << std::endl;
verify(interp_value_cpu, interp_value_gpu, result_size);
free(interp_value_cpu);
free(interp_value_gpu);
free(grid_values);
free(node_map);
free(weights);
}
void cuda_interpolation_function(fl_type* interp_value_gpu, int result_size, fl_type * grid_values, int grid_values_size, fl_type* weights, pos_type* node_map, int total_action_number, int interp_dim, int n_sim) {
int fl_size = sizeof(fl_type);
int pos_size = sizeof(pos_type);
auto start = std::chrono::steady_clock::now();
//device versions of the inputs
fl_type * grid_values_device;
fl_type* weights_device;
pos_type * node_map_device;
fl_type *interp_value_device;
int lenght_node_map = interp_dim*total_action_number;
std::cout << "size grid_values: " << grid_values_size <<std::endl;
std::cout << "size weights: " << lenght_node_map << std::endl;
std::cout << "size interp_value: " << result_size << std::endl;
//allocating and moving to the GPU the inputs
auto error_code=cudaMalloc((void**)&grid_values_device, grid_values_size*fl_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of the grid_values" << std::endl;
}
error_code=cudaMemcpy(grid_values_device, grid_values, grid_values_size*fl_size, cudaMemcpyHostToDevice);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMemcpy of the grid_values" << std::endl;
}
error_code=cudaMalloc((void**)&weights_device, lenght_node_map*fl_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of the weights" << std::endl;
}
error_code=cudaMemcpy(weights_device, weights, lenght_node_map*fl_size, cudaMemcpyHostToDevice);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMemcpy of the weights" << std::endl;
}
error_code=cudaMalloc((void**)&node_map_device, lenght_node_map*pos_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of node_map" << std::endl;
}
error_code=cudaMemcpy(node_map_device, node_map, lenght_node_map*pos_size, cudaMemcpyHostToDevice);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMemcpy of node_map" << std::endl;
}
error_code=cudaMalloc((void**)&interp_value_device, result_size*fl_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of interp_value_device " << std::endl;
}
auto elapsed_moving = std::chrono::steady_clock::now() - start;
float ms_moving = std::chrono::duration_cast<ms>(elapsed_moving).count();
cudaDeviceSynchronize();
//1d
int block_size = 1024;
int num_blocks = (result_size + block_size - 1) / block_size;
std::cout << "num_blocks:" << num_blocks << std::endl;
interp_kernel << < num_blocks, block_size >> > (interp_value_device, weights_device, node_map_device, grid_values_device, interp_dim, n_sim, total_action_number);
//2d
//int block_size2 = 32; //each block will have block_size2*block_size2 threads
//dim3 num_blocks2(block_size2, block_size2);
//int x_grid = (total_action_number + block_size2 - 1) / block_size2;
//int y_grid = (n_sim + block_size2 - 1) / block_size2;
//dim3 grid_size2(x_grid, y_grid);
//std::cout <<"grid:"<< x_grid<<" x "<< y_grid<<std::endl;
//interp_kernel2D <<< grid_size2, num_blocks2 >>> (interp_value_device, weights_device, node_map_device, grid_values_device, interp_dim, n_sim, total_action_number);
cudaDeviceSynchronize();
cudaError err = cudaGetLastError();
if (cudaSuccess != err)
{
std::cout << "Cuda kernel failed! " << cudaGetErrorString(err) <<std::endl;
}
start = std::chrono::steady_clock::now();
cudaMemcpy(interp_value_gpu, interp_value_device, result_size*fl_size, cudaMemcpyDeviceToHost);
auto elapsed_moving_back = std::chrono::steady_clock::now() - start;
float ms_moving_back = std::chrono::duration_cast<ms>(elapsed_moving_back).count();
std::cout << "Time spent moving the data to the GPU:" << ms_moving << " ms"<<std::endl;
std::cout << "Time spent moving the results back to the host: " << ms_moving_back << " ms" << std::endl;
cudaFree(interp_value_device);
cudaFree(weights_device);
cudaFree(node_map_device);
cudaFree(grid_values_device);
}
而且,我將非常感謝任何關於如何提高代碼性能的指導。
任何時候你有麻煩了CUDA代碼,我建議做適當的CUDA錯誤檢查(你大多似乎在做), 並運行你的代碼的時間cuda-memcheck
。 最后一個實用程序與Nsight VSE中的“啟用內存檢查器”相似,但並不完全相同。 但是,Nsight VSE內存檢查器可能已經給出了相同的指示。
在C(或C ++)中,數組的索引通常從0開始。因此,要測試越界索引,我必須檢查生成的索引是否等於或大於數組的大小。 但是在您的情況下,您僅測試以下項目:
if (row > n_sim || column > num_cols) return;
您在1D內核和2D內核中都犯了類似的錯誤,盡管您認為1D內核可以正常工作,但實際上是在進行越界訪問。 如果您使用上述的cuda-memcheck
實用程序(或可能還可以使用可以在Nsight VSE中啟用的內存檢查器)運行,則可以驗證這一點。
當我在pastebin鏈接中修改您的代碼以使用正確的范圍/邊界檢查時, cuda-memcheck
報告沒有錯誤,並且您的程序報告了正確的結果。 我已經測試了這兩種情況,但是下面的代碼已從您的pastebin鏈接修改為取消注釋2D情況,並使用它代替了1D情況:
$ cat t375.cu | more
#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>
#include <assert.h>
#include <iostream>
#include <random>
#include <chrono>
typedef float fl_type;
typedef int pos_type;
typedef std::chrono::milliseconds ms;
//declaration of the cuda function
void cuda_interpolation_function(fl_type* interp_value_back, int result_size, fl
_type * grid_values, int grid_values_size, fl_type* weights, pos_type* node_map,
int total_action_number, int interp_dim, int n_sim);
fl_type iterp_cpu(fl_type* weights, pos_type* node_map, fl_type* grid_values, in
t& row, int& column, int& interp_dim, int& n_sim) {
int w_p = column*interp_dim;
fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
res += weights[w_p + inter_point] * grid_values[node_map[w_p + inter_poi
nt] * n_sim + row];
}
return res;
}
__global__ void interp_kernel(fl_type * d_matrix, fl_type* weights, pos_type* no
de_map, fl_type* grid_values, int interp_dim, int n_sim, int num_cols) {
int index = blockIdx.x * blockDim.x + threadIdx.x;
int column = index / n_sim;
int row = index % n_sim;
int w_p = column*interp_dim;
if (row >= n_sim || column >= num_cols) return; // modified
fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
res += weights[w_p + inter_point] * grid_values[row + node_map[w_p + int
er_point] * n_sim];
}
d_matrix[index] = res;
}
__global__ void interp_kernel2D(fl_type * d_matrix, fl_type* weights, pos_type*
node_map, fl_type* grid_values, int interp_dim, int n_sim, int num_cols) {
int column = blockIdx.x * blockDim.x + threadIdx.x;
int row = blockIdx.y * blockDim.y + threadIdx.y;
int index = column*n_sim + row;
int w_p = column*interp_dim;
if (row >= n_sim || column >= num_cols) return; // modified
fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
res += weights[w_p + inter_point] * grid_values[row + node_map[w_p + int
er_point] * n_sim];
}
d_matrix[index] = res;
}
void verify(fl_type *host, fl_type *device, int size) {
int count = 0;
int count_zero = 0;
for (int i = 0; i < size; i++) {
if (host[i] != device[i]) {
count++;
//std::cout <<"pos: " <<i<< " CPU:" <<h[i] << ", GPU: " << d[
i] <<std::endl;
assert(host[i] == device[i]);
if (device[i] == 0.0)
count_zero++;
}
}
if (count) {
std::cout << "Non matching: " << count << "out of " << size << "(" << (f
loat(count) / size * 100) << "%)" << std::endl;
std::cout << "Zeros returned from the device: " << count_zero <<"(" << (
float(count_zero) / size * 100) << "%)" << std::endl;
}
else
std::cout << "Perfect match!" << std::endl;
}
int main() {
int fl_size = sizeof(fl_type);
int pos_size = sizeof(pos_type);
int dim = 5; // range: 2-5
int number_nodes = 5500; // range: 10.000-500.000
int max_actions = 12; // range: 6-200
int n_sim = 1000; // range: 1.000-10.000
int interp_dim = std::pow(2, dim);
int grid_values_size = n_sim*number_nodes;
std::default_random_engine generator;
std::normal_distribution<fl_type> normal_dist(0.0, 1);
std::uniform_int_distribution<> uniform_dist(0, number_nodes - 1);
double bit_allocated = 0;
fl_type * grid_values; //flattened 2d array, containing the value of the grid (n_sims x number_nodes)
grid_values = (fl_type *)malloc(grid_values_size * fl_size);
bit_allocated += grid_values_size * fl_size;
for (int i = 0; i < grid_values_size; i++)
grid_values[i] = normal_dist(generator);
pos_type * map_node2values_start; //vector that maps each node to the first column of the result matrix regarding that done
pos_type * map_node2values_how_many; //vector that stores how many action we have per node
map_node2values_start = (pos_type *)malloc(number_nodes * pos_size);
map_node2values_how_many = (pos_type *)malloc(number_nodes * pos_size);
bit_allocated += 2 * (number_nodes * pos_size);
for (int i = 0; i < number_nodes; i++) {
//each node as simply max_actions
map_node2values_start[i] = max_actions*i;
map_node2values_how_many[i] = max_actions;
}
//total number of actions, which is amount of column of the results
int total_action_number = map_node2values_start[number_nodes - 1] + map_node2values_how_many[number_nodes - 1];
//vector that keep tracks of the columnt to grab, and their weight in the interpolation
fl_type* weights;
pos_type * node_map;
weights = (fl_type *)malloc(total_action_number*interp_dim * pos_size);
bit_allocated += total_action_number * fl_size;
node_map = (pos_type *)malloc(total_action_number*interp_dim * pos_size);
bit_allocated += total_action_number * pos_size;
//filling with random numbers
for (int i = 0; i < total_action_number*interp_dim; i++) {
node_map[i] = uniform_dist(generator); // picking random column
weights[i] = 1.0 / interp_dim; // uniform weights
}
std::cout << "done filling!" << std::endl;
std::cout << bit_allocated / 8 / 1024 / 1024 << "MB allocated" << std::endl;
int result_size = n_sim*total_action_number;
fl_type *interp_value_cpu;
bit_allocated += result_size* fl_size;
interp_value_cpu = (fl_type *)malloc(result_size* fl_size);
auto start = std::chrono::steady_clock::now();
for (int row = 0; row < n_sim; row++) {
for (int column = 0; column < total_action_number; column++) {
auto zz = iterp_cpu(weights, node_map, grid_values, row, column, interp_dim, n_sim);
interp_value_cpu[column*n_sim + row] = zz;
}
}
auto elapsed_cpu = std::chrono::steady_clock::now() - start;
std::cout << "Crunching values on the CPU (serial): " << std::chrono::duration_cast<ms>(elapsed_cpu).count() / 1000.0 << "s" << std::endl;
int * pp;
cudaMalloc((void**)&pp, sizeof(int)); //initializing the device, to not affect the benchmark
fl_type *interp_value_gpu;
interp_value_gpu = (fl_type *)malloc(result_size* fl_size);
start = std::chrono::steady_clock::now();
cuda_interpolation_function(interp_value_gpu, result_size, grid_values, grid_values_size, weights, node_map, total_action_number, interp_dim, n_sim);
auto elapsed_gpu = std::chrono::steady_clock::now() - start;
std::cout << "Crunching values on the GPU: " << std::chrono::duration_cast<ms>(elapsed_gpu).count() / 1000.0 << "s" << std::endl;
float ms_cpu = std::chrono::duration_cast<ms>(elapsed_cpu).count();
float ms_gpu = std::chrono::duration_cast<ms>(elapsed_gpu).count();
int n_proc = 4;
std::cout << "Performance: " << (ms_gpu- ms_cpu / n_proc) / (ms_cpu / n_proc) * 100 << " % less time than parallel CPU!" << std::endl;
verify(interp_value_cpu, interp_value_gpu, result_size);
free(interp_value_cpu);
free(interp_value_gpu);
free(grid_values);
free(node_map);
free(weights);
}
void cuda_interpolation_function(fl_type* interp_value_gpu, int result_size, fl_type * grid_values, int grid_values_size, fl_type* weights, pos_type* node_map, int total_action_number, int interp_dim, int n_sim) {
int fl_size = sizeof(fl_type);
int pos_size = sizeof(pos_type);
auto start = std::chrono::steady_clock::now();
//device versions of the inputs
fl_type * grid_values_device;
fl_type* weights_device;
pos_type * node_map_device;
fl_type *interp_value_device;
int lenght_node_map = interp_dim*total_action_number;
std::cout << "size grid_values: " << grid_values_size <<std::endl;
std::cout << "size weights: " << lenght_node_map << std::endl;
std::cout << "size interp_value: " << result_size << std::endl;
//allocating and moving to the GPU the inputs
auto error_code=cudaMalloc((void**)&grid_values_device, grid_values_size*fl_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of the grid_values" << std::endl;
}
error_code=cudaMemcpy(grid_values_device, grid_values, grid_values_size*fl_size, cudaMemcpyHostToDevice);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMemcpy of the grid_values" << std::endl;
}
error_code=cudaMalloc((void**)&weights_device, lenght_node_map*fl_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of the weights" << std::endl;
}
error_code=cudaMemcpy(weights_device, weights, lenght_node_map*fl_size, cudaMemcpyHostToDevice);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMemcpy of the weights" << std::endl;
}
error_code=cudaMalloc((void**)&node_map_device, lenght_node_map*pos_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of node_map" << std::endl;
}
error_code=cudaMemcpy(node_map_device, node_map, lenght_node_map*pos_size, cudaMemcpyHostToDevice);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMemcpy of node_map" << std::endl;
}
error_code=cudaMalloc((void**)&interp_value_device, result_size*fl_size);
if (error_code != cudaSuccess) {
std::cout << "Error during cudaMalloc of interp_value_device " << std::endl;
}
auto elapsed_moving = std::chrono::steady_clock::now() - start;
float ms_moving = std::chrono::duration_cast<ms>(elapsed_moving).count();
cudaDeviceSynchronize();
//1d
#if 0
int block_size = 1024;
int num_blocks = (result_size + block_size - 1) / block_size;
std::cout << "num_blocks:" << num_blocks << std::endl;
interp_kernel << < num_blocks, block_size >> > (interp_value_device, weights_device, node_map_device, grid_values_device, interp_dim, n_sim, total_action_number);
#endif
//2d
int block_size2 = 32; //each block will have block_size2*block_size2 threads
dim3 num_blocks2(block_size2, block_size2);
int x_grid = (total_action_number + block_size2 - 1) / block_size2;
int y_grid = (n_sim + block_size2 - 1) / block_size2;
dim3 grid_size2(x_grid, y_grid);
std::cout <<"grid:"<< x_grid<<" x "<< y_grid<<std::endl;
interp_kernel2D <<< grid_size2, num_blocks2 >>> (interp_value_device, weights_device, node_map_device, grid_values_device, interp_dim, n_sim, total_action_number);
cudaDeviceSynchronize();
cudaError err = cudaGetLastError();
if (cudaSuccess != err)
{
std::cout << "Cuda kernel failed! " << cudaGetErrorString(err) <<std::endl;
}
start = std::chrono::steady_clock::now();
cudaMemcpy(interp_value_gpu, interp_value_device, result_size*fl_size, cudaMemcpyDeviceToHost);
auto elapsed_moving_back = std::chrono::steady_clock::now() - start;
float ms_moving_back = std::chrono::duration_cast<ms>(elapsed_moving_back).count();
std::cout << "Time spent moving the data to the GPU:" << ms_moving << " ms"<<std::endl;
std::cout << "Time spent moving the results back to the host: " << ms_moving_back << " ms" << std::endl;
cudaFree(interp_value_device);
cudaFree(weights_device);
cudaFree(node_map_device);
cudaFree(grid_values_device);
}
$ nvcc -arch=sm_52 -o t375 t375.cu -std=c++11
$ cuda-memcheck ./t375
========= CUDA-MEMCHECK
done filling!
2.69079MB allocated
Crunching values on the CPU (serial): 30.081s
size grid_values: 5500000
size weights: 2112000
size interp_value: 66000000
grid:2063 x 32
Time spent moving the data to the GPU:31 ms
Time spent moving the results back to the host: 335 ms
Crunching values on the GPU: 7.089s
Performance: -5.73452 % less time than parallel CPU!
Perfect match!
========= ERROR SUMMARY: 0 errors
$
請注意, cuda-memcheck
會減慢您的程序在GPU上的執行速度,以進行嚴格的內存邊界檢查。 因此,性能可能與普通情況不符。 這是“常規”運行的樣子:
$ ./t375
done filling!
2.69079MB allocated
Crunching values on the CPU (serial): 30.273s
size grid_values: 5500000
size weights: 2112000
size interp_value: 66000000
grid:2063 x 32
Time spent moving the data to the GPU:32 ms
Time spent moving the results back to the host: 332 ms
Crunching values on the GPU: 1.161s
Performance: -84.6596 % less time than parallel CPU!
Perfect match!
$
您正在訪問分配的塊之外的內存。 要檢查行索引和列索引是否在范圍內:
if (row >= n_rows || column >= num_cols) return; // Do this
if (row > n_rows || column > num_cols) return; // Instead of this
在平面版本中,此int row = index % n_rows;
使row
保持在n_rows
以下。 您只能訪問已分配內存之外的一列,對於較小的矩陣,該列可能仍與內存對齊有關。 Python演示 。
第二個版本確實訪問了一個額外的列plus和extra元素, 以及每行一個額外的元素(下一行的第一個元素),如下所示:
int row = blockIdx.y * blockDim.y + threadIdx.y;
行索引不再保持在有效范圍內。 Python演示 。
查看您的pastebin,可能是它破裂的地方:
44. fl_type res = weights[w_p] * grid_values[row + node_map[w_p] * n_sim];
^^^
45. for (int inter_point = 1; inter_point < interp_dim; inter_point++) {
46. res += weights[w_p + inter_point] * \
grid_values[row + node_map[w_p + inter_point] * n_sim];
^^^
47. }
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.