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

New optimized NMS

Browse Source
This commit is contained in:
Marcos Luciano
2022-08-15 02:48:23 -03:00
parent 8ad99d3f8d
commit c8a4a49f16
24 changed files with 206 additions and 394 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
nvdsinfer_custom_impl_Yolo/yoloForward_nc.cu
View File

@@ -6,7 +6,7 @@
#include <stdint.h>
__global__ void gpuYoloLayer_nc(
const float* input, int* d_indexes, float* d_scores, float* d_boxes, int* d_classes, int* countData,
const float* input, int* num_detections, float* detection_boxes, float* detection_scores, int* detection_classes,
const float scoreThreshold, const uint netWidth, const uint netHeight, const uint gridSizeX, const uint gridSizeY,
const uint numOutputClasses, const uint numBBoxes, const float scaleXY, const float* anchors, const int* mask)
{
@@ -26,7 +26,7 @@ __global__ void gpuYoloLayer_nc(
if (objectness < scoreThreshold)
return;
int count = (int)atomicAdd(countData, 1);
int count = (int)atomicAdd(num_detections, 1);
const float alpha = scaleXY;
const float beta = -0.5 * (scaleXY - 1);
@@ -62,23 +62,22 @@ __global__ void gpuYoloLayer_nc(
}
}
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;
detection_boxes[count * 4 + 0] = x - 0.5 * w;
detection_boxes[count * 4 + 1] = y - 0.5 * h;
detection_boxes[count * 4 + 2] = x + 0.5 * w;
detection_boxes[count * 4 + 3] = y + 0.5 * h;
detection_scores[count] = objectness * maxProb;
detection_classes[count] = maxIndex;
}
cudaError_t cudaYoloLayer_nc(
const void* input, void* d_indexes, void* d_scores, void* d_boxes, void* d_classes, void* countData,
const void* input, void* num_detections, void* detection_boxes, void* detection_scores, void* detection_classes,
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 float& scaleXY, const void* anchors, const void* mask, cudaStream_t stream);
cudaError_t cudaYoloLayer_nc(
const void* input, void* d_indexes, void* d_scores, void* d_boxes, void* d_classes, void* countData,
const void* input, void* num_detections, void* detection_boxes, void* detection_scores, void* detection_classes,
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 float& scaleXY, const void* anchors, const void* mask, cudaStream_t stream)
@@ -92,10 +91,10 @@ cudaError_t cudaYoloLayer_nc(
{
gpuYoloLayer_nc<<<number_of_blocks, threads_per_block, 0, stream>>>(
reinterpret_cast<const float*>(input) + (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),
reinterpret_cast<int*>(num_detections) + (batch),
reinterpret_cast<float*>(detection_boxes) + (batch * 4 * outputSize),
reinterpret_cast<float*>(detection_scores) + (batch * outputSize),
reinterpret_cast<int*>(detection_classes) + (batch * outputSize),
scoreThreshold, netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY,
reinterpret_cast<const float*>(anchors), reinterpret_cast<const int*>(mask));
}