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

Minor fixes

Browse Source
This commit is contained in:
Marcos Luciano
2022-02-21 23:46:29 -03:00
parent 66962cfeb8
commit 555152064e
29 changed files with 416 additions and 541 deletions
Download Patch File Download Diff File Expand all files Collapse all files

45
nvdsinfer_custom_impl_Yolo/yoloForward_v2.cu
View File

@@ -11,8 +11,9 @@
inline __device__ float sigmoidGPU(const float& x) { return 1.0f / (1.0f + __expf(-x)); }
__device__ void softmaxGPU(const float* input, const int bbindex, const int numGridCells,
uint z_id, const uint numOutputClasses, float temp, float* output)
__device__ void softmaxGPU(
const float* input, const int bbindex, const int numGridCells, uint z_id,
const uint numOutputClasses, float temp, float* output)
{
int i;
float sum = 0;
@@ -31,8 +32,10 @@ __device__ void softmaxGPU(const float* input, const int bbindex, const int numG
}
}
__global__ void gpuRegionLayer(const float* input, float* output, float* softmax, const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses,
const uint numBBoxes, const float* anchors)
__global__ void gpuRegionLayer(
const float* input, float* output, float* softmax, const uint netWidth, const uint netHeight,
const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses, const uint numBBoxes,
const float* anchors)
{
uint x_id = blockIdx.x * blockDim.x + threadIdx.x;
uint y_id = blockIdx.y * blockDim.y + threadIdx.y;
@@ -47,16 +50,20 @@ __global__ void gpuRegionLayer(const float* input, float* output, float* softmax
const int bbindex = y_id * gridSizeX + x_id;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]) + x_id;
= (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)])
+ x_id) * netWidth / gridSizeX;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]) + y_id;
= (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)])
+ y_id) * netHeight / gridSizeY;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]) * anchors[z_id * 2];
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)])
* anchors[z_id * 2] * netWidth / gridSizeX;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]) * anchors[z_id * 2 + 1];
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)])
* anchors[z_id * 2 + 1] * netHeight / gridSizeY;
softmaxGPU(input, bbindex, numGridCells, z_id, numOutputClasses, 1.0, softmax);
@@ -85,25 +92,29 @@ __global__ void gpuRegionLayer(const float* input, float* output, float* softmax
= maxIndex;
}
cudaError_t cudaYoloLayer_v2(const void* input, void* output, void* softmax, const uint& batchSize, const uint& gridSizeX, const uint& gridSizeY,
const uint& numOutputClasses, const uint& numBBoxes, uint64_t outputSize, cudaStream_t stream,
const void* anchors);
cudaError_t cudaRegionLayer(
const void* input, void* output, void* softmax, const uint& batchSize, const uint& netWidth,
const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses,
const uint& numBBoxes, uint64_t& outputSize, const void* anchors, cudaStream_t stream);
cudaError_t cudaYoloLayer_v2(const void* input, void* output, void* softmax, const uint& batchSize, const uint& gridSizeX, const uint& gridSizeY,
const uint& numOutputClasses, const uint& numBBoxes, uint64_t outputSize, cudaStream_t stream,
const void* anchors)
cudaError_t cudaRegionLayer(
const void* input, void* output, void* softmax, const uint& batchSize, const uint& netWidth,
const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses,
const uint& numBBoxes, uint64_t& outputSize, const void* anchors, cudaStream_t stream)
{
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);
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),
reinterpret_cast<float*>(softmax) + (batch * outputSize), gridSizeX, gridSizeY, numOutputClasses,
numBBoxes, reinterpret_cast<const float*>(anchors));
reinterpret_cast<float*>(output) + (batch * outputSize),
reinterpret_cast<float*>(softmax) + (batch * outputSize),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes,
reinterpret_cast<const float*>(anchors));
}
return cudaGetLastError();
}