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

38
nvdsinfer_custom_impl_Yolo/yoloForward_nc.cu
View File

@@ -9,8 +9,10 @@
#include <stdio.h>
#include <string.h>
__global__ void gpuYoloLayer_nc(const float* input, float* output, const uint gridSizeX, const uint gridSizeY, const uint numOutputClasses,
const uint numBBoxes, const float scaleXY, const float* anchors, const int* mask)
__global__ void gpuYoloLayer_nc(
const float* input, float* output, 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)
{
uint x_id = blockIdx.x * blockDim.x + threadIdx.x;
uint y_id = blockIdx.y * blockDim.y + threadIdx.y;
@@ -28,16 +30,20 @@ __global__ void gpuYoloLayer_nc(const float* input, float* output, const uint gr
const float beta = -0.5 * (scaleXY - 1);
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)] * alpha + beta + x_id;
= (input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 0)]
* alpha + beta + x_id) * netWidth / gridSizeX;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)] * alpha + beta + y_id;
= (input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 1)]
* alpha + beta + y_id) * netHeight / gridSizeY;
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)]
= __powf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)] * 2, 2) * anchors[mask[z_id] * 2];
= __powf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 2)] * 2, 2)
* anchors[mask[z_id] * 2];
output[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)]
= __powf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)] * 2, 2) * anchors[mask[z_id] * 2 + 1];
= __powf(input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 3)] * 2, 2)
* anchors[mask[z_id] * 2 + 1];
const float objectness
= input[bbindex + numGridCells * (z_id * (5 + numOutputClasses) + 4)];
@@ -64,24 +70,28 @@ __global__ void gpuYoloLayer_nc(const float* input, float* output, const uint gr
= maxIndex;
}
cudaError_t cudaYoloLayer_nc(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 float scaleXY, const void* anchors, const void* mask);
cudaError_t cudaYoloLayer_nc(
const void* input, void* output, 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 float& scaleXY, const void* anchors, const void* mask, cudaStream_t stream);
cudaError_t cudaYoloLayer_nc(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 float scaleXY, const void* anchors, const void* mask)
cudaError_t cudaYoloLayer_nc(
const void* input, void* output, 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 float& scaleXY, const void* anchors, const void* mask, 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)
{
gpuYoloLayer_nc<<<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, scaleXY, reinterpret_cast<const float*>(anchors), reinterpret_cast<const int*>(mask));
reinterpret_cast<float*>(output) + (batch * outputSize),
netWidth, netHeight, gridSizeX, gridSizeY, numOutputClasses, numBBoxes, scaleXY,
reinterpret_cast<const float*>(anchors), reinterpret_cast<const int*>(mask));
}
return cudaGetLastError();
}