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

GPU Batched NMS

Browse Source
This commit is contained in:
Marcos Luciano
2022-06-19 03:25:50 -03:00
parent 2300e3b44b
commit f621c0f429
24 changed files with 835 additions and 654 deletions
Download Patch File Download Diff File Expand all files Collapse all files

69
nvdsinfer_custom_impl_Yolo/yoloForward_v2.cu
View File

@@ -3,17 +3,13 @@
* https://www.github.com/marcoslucianops
*/
#include <cuda.h>
#include <cuda_runtime.h>
#include <stdint.h>
#include <stdio.h>
#include <string.h>
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)
const float* input, const int bbindex, const int numGridCells, uint z_id, const uint numOutputClasses, float temp,
float* output)
{
int i;
float sum = 0;
@@ -33,9 +29,9 @@ __device__ void softmaxGPU(
}
__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)
const float* input, float* softmax, int* d_indexes, float* d_scores, float* d_boxes, int* d_classes, int* countData,
const float scoreThreshold, 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;
@@ -49,27 +45,31 @@ __global__ void gpuRegionLayer(
const int numGridCells = gridSizeX * gridSizeY;
const int bbindex = y_id * gridSizeX + x_id;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
const float objectness
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]);
if (objectness < scoreThreshold) return;
int count = (int)atomicAdd(&countData[0], 1);
float x
= (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)])
+ x_id) * netWidth / gridSizeX;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
float y
= (sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)])
+ y_id) * netHeight / gridSizeY;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
float w
= __expf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)])
* anchors[z_id * 2] * netWidth / gridSizeX;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]
float h
= __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);
const float objectness
= sigmoidGPU(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]);
float maxProb = 0.0f;
int maxIndex = -1;
@@ -85,22 +85,26 @@ __global__ void gpuRegionLayer(
}
}
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)]
= objectness * maxProb;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 5)]
= maxIndex;
d_indexes[count] = count;
d_scores[count] = objectness * maxProb + 1.f;
d_boxes[count * 4 + 0] = x - 0.5 * w;
d_boxes[count * 4 + 1] = y - 0.5 * h;
d_boxes[count * 4 + 2] = x + 0.5 * w;
d_boxes[count * 4 + 3] = y + 0.5 * h;
d_classes[count] = maxIndex;
}
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);
const void* input, void* softmax, void* d_indexes, void* d_scores, void* d_boxes, void* d_classes, void* countData,
const uint& batchSize, uint64_t& inputSize, uint64_t& outputSize, const float& scoreThreshold, const uint& netWidth,
const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses, const uint& numBBoxes,
const void* anchors, cudaStream_t stream);
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)
const void* input, void* softmax, void* d_indexes, void* d_scores, void* d_boxes, void* d_classes, void* countData,
const uint& batchSize, uint64_t& inputSize, uint64_t& outputSize, const float& scoreThreshold, const uint& netWidth,
const uint& netHeight, const uint& gridSizeX, const uint& gridSizeY, const uint& numOutputClasses, const uint& numBBoxes,
const void* anchors, cudaStream_t stream)
{
dim3 threads_per_block(16, 16, 4);
dim3 number_of_blocks((gridSizeX / threads_per_block.x) + 1,
@@ -110,10 +114,13 @@ cudaError_t cudaRegionLayer(
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),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes,
reinterpret_cast<const float*>(input) + (batch * inputSize),
reinterpret_cast<float*>(softmax) + (batch * inputSize),
reinterpret_cast<int*>(d_indexes) + (batch * outputSize),
reinterpret_cast<float*>(d_scores) + (batch * outputSize),
reinterpret_cast<float*>(d_boxes) + (batch * 4 * outputSize),
reinterpret_cast<int*>(d_classes) + (batch * outputSize), reinterpret_cast<int*>(countData) + (batch),
scoreThreshold, netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes,
reinterpret_cast<const float*>(anchors));
}
return cudaGetLastError();