augustin64/tipe
1
1
Fork 0
mirror of https://github.com/augustin64/projet-tipe synced 2025-01-23 15:16:26 +01:00
Code Issues Packages Projects Releases Wiki Activity

cnn/convolution: fix misaligned address issue

Browse Source
This commit is contained in:
augustin64 2023-05-15 11:55:50 +02:00
parent 2f333bfc1d
commit 37ba3a5976
3 changed files with 21 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
src/cnn/convolution.c
View File

@ -42,27 +42,27 @@ void make_convolution_cpu(Kernel_cnn* kernel, float*** input, float*** output, i
#ifdef __CUDACC__ #ifdef __CUDACC__
__global__ void make_convolution_kernel(Kernel_cnn* kernel, float*** input, float*** output, int output_width, int stride, int padding) { __global__ void make_convolution_kernel(float**** weights, float*** bias, int k_size, int rows, int columns, float*** input, float*** output, int output_width, int stride, int padding) {
// Équivalents respectifs de i, j et k dans la boucle effectuée par le cpu // Équivalents respectifs de i, j et k dans la boucle effectuée par le cpu
int idx = threadIdx.x + blockDim.x*blockIdx.x; // < kernel->columns int idx = threadIdx.x + blockDim.x*blockIdx.x; // < kernel->columns
int idy = threadIdx.y + blockDim.y*blockIdx.y; // < min(output_width, k_size) int idy = threadIdx.y + blockDim.y*blockIdx.y; // < min(output_width, k_size)
int idz = threadIdx.z + blockDim.z*blockIdx.z; // < min(output_width, k_size) int idz = threadIdx.z + blockDim.z*blockIdx.z; // < min(output_width, k_size)
int max_move = kernel->k_size - padding; int max_move = k_size - padding;
int input_width = output_width*stride - 2*padding + kernel->k_size - stride; int input_width = output_width*stride - 2*padding + k_size - stride;
if (idx >= kernel->columns || idy >= output_width || idz >= output_width) { if (idx >= columns || idy >= output_width || idz >= output_width) {
return; return;
} }
float f = kernel->bias[idx][idy][idz]; float f = bias[idx][idy][idz];
for (int a=0; a < kernel->rows; a++) { for (int a=0; a < rows; a++) {
for (int b=-padding; b < max_move; b++) { for (int b=-padding; b < max_move; b++) {
for (int c=-padding; c < max_move; c++) { for (int c=-padding; c < max_move; c++) {
int idy_2 = idy*stride+b; int idy_2 = idy*stride+b;
int idz_2 = idz*stride+c; int idz_2 = idz*stride+c;
if (not_outside(idy_2, idz_2, 0, input_width)) { if (not_outside(idy_2, idz_2, 0, input_width)) {
f += kernel->weights[a][idx][b][c]*input[a][idy_2][idz_2]; f += weights[a][idx][b][c]*input[a][idy_2][idz_2];
} }
} }
} }
@ -76,7 +76,9 @@ void make_convolution_device(Kernel_cnn* kernel, float*** input, float*** output
dim3 gridSize(i_div_up(kernel->columns, BLOCKSIZE_x), i_div_up(output_width, BLOCKSIZE_y), i_div_up(output_width, BLOCKSIZE_z)); dim3 gridSize(i_div_up(kernel->columns, BLOCKSIZE_x), i_div_up(output_width, BLOCKSIZE_y), i_div_up(output_width, BLOCKSIZE_z));
dim3 blockSize(BLOCKSIZE_x, BLOCKSIZE_y, BLOCKSIZE_z); dim3 blockSize(BLOCKSIZE_x, BLOCKSIZE_y, BLOCKSIZE_z);
make_convolution_kernel<<<gridSize, blockSize>>>(kernel, input, output, output_width, stride, padding); // We can't pass `kernel` directly to the CUDA kernel function
// as it will create a 'misaligned adress' error
make_convolution_kernel<<<gridSize, blockSize>>>(kernel->weights, kernel->bias, kernel->k_size, kernel->rows, kernel->columns, input, output, output_width, stride, padding);
gpuErrchk( cudaPeekAtLastError() ); gpuErrchk( cudaPeekAtLastError() );
gpuErrchk( cudaDeviceSynchronize() ); gpuErrchk( cudaDeviceSynchronize() );
} }

18
src/cnn/convolution.cu
View File

@ -42,27 +42,27 @@ void make_convolution_cpu(Kernel_cnn* kernel, float*** input, float*** output, i
#ifdef __CUDACC__ #ifdef __CUDACC__
__global__ void make_convolution_kernel(Kernel_cnn* kernel, float*** input, float*** output, int output_width, int stride, int padding) { __global__ void make_convolution_kernel(float**** weights, float*** bias, int k_size, int rows, int columns, float*** input, float*** output, int output_width, int stride, int padding) {
// Équivalents respectifs de i, j et k dans la boucle effectuée par le cpu // Équivalents respectifs de i, j et k dans la boucle effectuée par le cpu
int idx = threadIdx.x + blockDim.x*blockIdx.x; // < kernel->columns int idx = threadIdx.x + blockDim.x*blockIdx.x; // < kernel->columns
int idy = threadIdx.y + blockDim.y*blockIdx.y; // < min(output_width, k_size) int idy = threadIdx.y + blockDim.y*blockIdx.y; // < min(output_width, k_size)
int idz = threadIdx.z + blockDim.z*blockIdx.z; // < min(output_width, k_size) int idz = threadIdx.z + blockDim.z*blockIdx.z; // < min(output_width, k_size)
int max_move = kernel->k_size - padding; int max_move = k_size - padding;
int input_width = output_width*stride - 2*padding + kernel->k_size - stride; int input_width = output_width*stride - 2*padding + k_size - stride;
if (idx >= kernel->columns || idy >= output_width || idz >= output_width) { if (idx >= columns || idy >= output_width || idz >= output_width) {
return; return;
} }
float f = kernel->bias[idx][idy][idz]; float f = bias[idx][idy][idz];
for (int a=0; a < kernel->rows; a++) { for (int a=0; a < rows; a++) {
for (int b=-padding; b < max_move; b++) { for (int b=-padding; b < max_move; b++) {
for (int c=-padding; c < max_move; c++) { for (int c=-padding; c < max_move; c++) {
int idy_2 = idy*stride+b; int idy_2 = idy*stride+b;
int idz_2 = idz*stride+c; int idz_2 = idz*stride+c;
if (not_outside(idy_2, idz_2, 0, input_width)) { if (not_outside(idy_2, idz_2, 0, input_width)) {
f += kernel->weights[a][idx][b][c]*input[a][idy_2][idz_2]; f += weights[a][idx][b][c]*input[a][idy_2][idz_2];
} }
} }
} }
@ -76,7 +76,9 @@ void make_convolution_device(Kernel_cnn* kernel, float*** input, float*** output
dim3 gridSize(i_div_up(kernel->columns, BLOCKSIZE_x), i_div_up(output_width, BLOCKSIZE_y), i_div_up(output_width, BLOCKSIZE_z)); dim3 gridSize(i_div_up(kernel->columns, BLOCKSIZE_x), i_div_up(output_width, BLOCKSIZE_y), i_div_up(output_width, BLOCKSIZE_z));
dim3 blockSize(BLOCKSIZE_x, BLOCKSIZE_y, BLOCKSIZE_z); dim3 blockSize(BLOCKSIZE_x, BLOCKSIZE_y, BLOCKSIZE_z);
make_convolution_kernel<<<gridSize, blockSize>>>(kernel, input, output, output_width, stride, padding); // We can't pass `kernel` directly to the CUDA kernel function
// as it will create a 'misaligned adress' error
make_convolution_kernel<<<gridSize, blockSize>>>(kernel->weights, kernel->bias, kernel->k_size, kernel->rows, kernel->columns, input, output, output_width, stride, padding);
gpuErrchk( cudaPeekAtLastError() ); gpuErrchk( cudaPeekAtLastError() );
gpuErrchk( cudaDeviceSynchronize() ); gpuErrchk( cudaDeviceSynchronize() );
} }

2
src/cnn/include/convolution.h
View File

@ -10,7 +10,7 @@ void make_convolution_cpu(Kernel_cnn* kernel, float*** input, float*** output, i
/* /*
* Kernel de la convolution sur carte graphique * Kernel de la convolution sur carte graphique
*/ */
__global__ void make_convolution_kernel(int k_size, int columns, int rows, float* bias, size_t pitch_bias, float**** weights, size_t pitch_weights, float*** input, size_t pitch_input, float*** output, size_t pitch_output, int output_width, int stride, int padding); __global__ void make_convolution_kernel(float**** weights, float*** bias, int k_size, int rows, int columns, float*** input, float*** output, int output_width, int stride, int padding);
/* /*
* Effectue la convolution naïvement sur la carte graphique * Effectue la convolution naïvement sur la carte graphique