thebears/deepstream_yolo
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix YOLO kernels

Browse Source 
- Fix YOLO kernels
- Update deprecated functions
This commit is contained in:
unknown
2021-12-12 09:58:23 -03:00
parent ce35e17334
commit 9565254551
11 changed files with 316 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files

143
nvdsinfer_custom_impl_Yolo/yoloForward.cu
View File

@@ -21,7 +21,7 @@
inline __device__ float sigmoidGPU(const float& x) { return 1.0f / (1.0f + __expf(-x)); }
__global__ void gpuYoloLayer(const float* input, float* output, const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses,
const uint numBBoxes, const uint new_coords, const float scale_x_y)
const uint numBBoxes, const float scale_x_y)
{
uint x_id = blockIdx.x * blockDim.x + threadIdx.x;
uint y_id = blockIdx.y * blockDim.y + threadIdx.y;
@@ -35,97 +35,14 @@ __global__ void gpuYoloLayer(const float* input, float* output, const uint gridS
const int numGridCells = gridSizeX * gridSizeY;
const int bbindex = y_id * gridSizeX + x_id;
float alpha = scale_x_y;
float beta = -0.5 * (scale_x_y - 1);
if (new_coords == 1) {
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)] * alpha + beta;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)] * alpha + beta;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
= pow(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)] * 2, 2);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]
= pow(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)] * 2, 2);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)];
for (uint i = 0; i < numOutputClasses; ++i)
{
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))];
}
}
else if (new_coords == 0 && scale_x_y != 1) { // YOLOR incorrect param
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]) * 2.0 - 0.5;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]) * 2.0 - 0.5;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
= pow(sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]) * 2, 2);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]
= pow(sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]) * 2, 2);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]);
for (uint i = 0; i < numOutputClasses; ++i)
{
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]);
}
}
else {
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]) * alpha + beta;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]) * alpha + beta;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]);
for (uint i = 0; i < numOutputClasses; ++i)
{
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]);
}
}
}
__global__ void gpuRegionLayer(const float* input, float* output, const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses,
const uint numBBoxes)
{
uint x_id = blockIdx.x * blockDim.x + threadIdx.x;
uint y_id = blockIdx.y * blockDim.y + threadIdx.y;
uint z_id = blockIdx.z * blockDim.z + threadIdx.z;
if ((x_id >= gridSizeX) || (y_id >= gridSizeY) || (z_id >= numBBoxes))
{
return;
}
const int numGridCells = gridSizeX * gridSizeY;
const int bbindex = y_id * gridSizeX + x_id;
const float alpha = scale_x_y;
const float beta = -0.5 * (scale_x_y - 1);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]);
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]) * alpha + beta;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]);
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]) * alpha + beta;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]);
@@ -136,53 +53,31 @@ __global__ void gpuRegionLayer(const float* input, float* output, const uint gri
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]);
float temp = 1.0;
int i;
float sum = 0;
float largest = -INFINITY;
for(i = 0; i < numOutputClasses; ++i){
int val = input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))];
largest = (val>largest) ? val : largest;
}
for(i = 0; i < numOutputClasses; ++i){
float e = exp(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))] / temp - largest / temp);
sum += e;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))] = e;
}
for(i = 0; i < numOutputClasses; ++i){
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))] /= sum;
for (uint i = 0; i < numOutputClasses; ++i)
{
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + (5 + i))]);
}
}
cudaError_t cudaYoloLayer(const void* input, void* output, const uint& batchSize, const uint& gridSizeX, const uint& gridSizeY,
const uint& numOutputClasses, const uint& numBBoxes,
uint64_t outputSize, cudaStream_t stream, const uint modelCoords, const float modelScale, const uint modelType);
const uint& numOutputClasses, const uint& numBBoxes, uint64_t outputSize, cudaStream_t stream,
const float modelScale);
cudaError_t cudaYoloLayer(const void* input, void* output, const uint& batchSize, const uint& gridSizeX, const uint& gridSizeY,
const uint& numOutputClasses, const uint& numBBoxes,
uint64_t outputSize, cudaStream_t stream, const uint modelCoords, const float modelScale, const uint modelType)
const uint& numOutputClasses, const uint& numBBoxes, uint64_t outputSize, cudaStream_t stream,
const float modelScale)
{
dim3 threads_per_block(16, 16, 4);
dim3 number_of_blocks((gridSizeX / threads_per_block.x) + 1,
(gridSizeY / threads_per_block.y) + 1,
(numBBoxes / threads_per_block.z) + 1);
if (modelType == 1) {
for (unsigned int batch = 0; batch < batchSize; ++batch)
{
gpuYoloLayer<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*>(input) + (batch * outputSize),
reinterpret_cast<float*>(output) + (batch * outputSize), gridSizeX, gridSizeY, numOutputClasses,
numBBoxes, modelCoords, modelScale);
}
}
else if (modelType == 0) {
for (unsigned int batch = 0; batch < batchSize; ++batch)
{
gpuRegionLayer<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*>(input) + (batch * outputSize),
reinterpret_cast<float*>(output) + (batch * outputSize), gridSizeX, gridSizeY, numOutputClasses,
numBBoxes);
}
for (unsigned int batch = 0; batch < batchSize; ++batch)
{
gpuYoloLayer<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*>(input) + (batch * outputSize),
reinterpret_cast<float*>(output) + (batch * outputSize), gridSizeX, gridSizeY, numOutputClasses,
numBBoxes, modelScale);
}
return cudaGetLastError();
}