ANSCENTER/ANSCORE
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Use CPU resize before upload to GPU to remove PCIe bottleneck

Browse Source
This commit is contained in:
Tuan Nghia Nguyen
2026-04-04 22:29:08 +11:00
parent e134ebdf15
commit 98681f4da6
15 changed files with 572 additions and 493 deletions
Download Patch File Download Diff File Expand all files Collapse all files

156
modules/ANSODEngine/ANSRTYOLO.cpp
View File

@@ -462,50 +462,46 @@ namespace ANSCENTER {
// Early-out if CUDA context is dead (sticky error from CUVID crash etc.)
if (!m_nv12Helper.isCudaContextHealthy(_logger, "ANSRTYOLO")) return {};
cv::cuda::Stream stream;
cv::cuda::GpuMat gpuImg;
// Resolve source Mat (handle grayscale → BGR on CPU first)
if (inputImage.channels() == 1) {
cv::Mat img3Channel;
cv::cvtColor(inputImage, img3Channel, cv::COLOR_GRAY2BGR);
gpuImg.upload(img3Channel, stream);
} else {
gpuImg.upload(inputImage, stream);
// --- CPU preprocessing: resize + BGR→RGB before GPU upload ---
// Reduces PCIe transfer from 25 MB (4K BGR) to 1.2 MB (640×640 RGB).
// With 12 AI tasks uploading concurrently, this eliminates the WDDM
// SRW lock convoy that causes 400-580ms preprocess spikes.
cv::Mat srcImg = inputImage;
if (srcImg.channels() == 1) {
cv::cvtColor(srcImg, srcImg, cv::COLOR_GRAY2BGR);
}
// GPU: BGR → RGB
cv::cuda::GpuMat gpuRGB;
cv::cuda::cvtColor(gpuImg, gpuRGB, cv::COLOR_BGR2RGB, 0, stream);
outMeta.imgHeight = static_cast<float>(gpuRGB.rows);
outMeta.imgWidth = static_cast<float>(gpuRGB.cols);
outMeta.imgHeight = static_cast<float>(srcImg.rows);
outMeta.imgWidth = static_cast<float>(srcImg.cols);
if (outMeta.imgHeight > 0 && outMeta.imgWidth > 0) {
outMeta.ratio = 1.f / std::min(
inputDims[0].d[2] / static_cast<float>(gpuRGB.cols),
inputDims[0].d[1] / static_cast<float>(gpuRGB.rows));
inputDims[0].d[2] / static_cast<float>(srcImg.cols),
inputDims[0].d[1] / static_cast<float>(srcImg.rows));
// Check if model is classification (output ndims <= 2)
const auto& outputDims = m_trtEngine->getOutputDims();
const bool isClassification = !outputDims.empty() && outputDims[0].nbDims <= 2;
cv::cuda::GpuMat gpuResized;
if (gpuRGB.rows != inputH || gpuRGB.cols != inputW) {
// CPU resize to model input size
cv::Mat cpuResized;
if (srcImg.rows != inputH || srcImg.cols != inputW) {
if (isClassification) {
// Classification: direct resize (no letterbox padding)
cv::cuda::resize(gpuRGB, gpuResized, cv::Size(inputW, inputH),
0, 0, cv::INTER_LINEAR, stream);
}
else {
// Detection/Seg/Pose/OBB: letterbox resize + right-bottom pad (on GPU)
gpuResized = Engine<float>::resizeKeepAspectRatioPadRightBottom(
gpuRGB, inputH, inputW);
cv::resize(srcImg, cpuResized, cv::Size(inputW, inputH), 0, 0, cv::INTER_LINEAR);
} else {
cpuResized = Engine<float>::cpuResizeKeepAspectRatioPadRightBottom(srcImg, inputH, inputW);
}
} else {
gpuResized = gpuRGB;
cpuResized = srcImg;
}
// CPU BGR → RGB
cv::Mat cpuRGB;
cv::cvtColor(cpuResized, cpuRGB, cv::COLOR_BGR2RGB);
// Upload small image to GPU (1.2 MB instead of 25 MB for 4K)
cv::cuda::Stream stream;
cv::cuda::GpuMat gpuResized;
gpuResized.upload(cpuRGB, stream);
stream.waitForCompletion();
std::vector<cv::cuda::GpuMat> input{ std::move(gpuResized) };
@@ -878,26 +874,18 @@ namespace ANSCENTER {
"Empty input image at index " + std::to_string(i), __FILE__, __LINE__);
return {};
}
cv::cuda::GpuMat img;
if (inputImage.channels() == 1) {
cv::Mat img3Channel;
cv::cvtColor(inputImage, img3Channel, cv::COLOR_GRAY2BGR);
img.upload(img3Channel, stream);
}
else if (inputImage.channels() == 3) {
img.upload(inputImage, stream);
}
else {
// CPU preprocessing: resize + BGR→RGB before GPU upload
cv::Mat srcImg = inputImage;
if (srcImg.channels() == 1) {
cv::cvtColor(srcImg, srcImg, cv::COLOR_GRAY2BGR);
} else if (srcImg.channels() != 3) {
_logger.LogError("ANSRTYOLO::PreprocessBatch",
"Unsupported channel count at index " + std::to_string(i), __FILE__, __LINE__);
return {};
}
cv::cuda::GpuMat imgRGB;
cv::cuda::cvtColor(img, imgRGB, cv::COLOR_BGR2RGB, 0, stream);
outMetadata.imgHeights[i] = imgRGB.rows;
outMetadata.imgWidths[i] = imgRGB.cols;
outMetadata.imgHeights[i] = srcImg.rows;
outMetadata.imgWidths[i] = srcImg.cols;
if (outMetadata.imgHeights[i] <= 0 || outMetadata.imgWidths[i] <= 0) {
_logger.LogError("ANSRTYOLO::PreprocessBatch",
"Invalid dimensions for image " + std::to_string(i), __FILE__, __LINE__);
@@ -907,23 +895,27 @@ namespace ANSCENTER {
const auto& outputDims = m_trtEngine->getOutputDims();
const bool isClassification = !outputDims.empty() && outputDims[0].nbDims <= 2;
const float scaleW = inputW / static_cast<float>(imgRGB.cols);
const float scaleH = inputH / static_cast<float>(imgRGB.rows);
const float scaleW = inputW / static_cast<float>(srcImg.cols);
const float scaleH = inputH / static_cast<float>(srcImg.rows);
outMetadata.ratios[i] = isClassification ? 1.f : 1.f / std::min(scaleW, scaleH);
cv::cuda::GpuMat resized;
if (imgRGB.rows != inputH || imgRGB.cols != inputW) {
cv::Mat cpuResized;
if (srcImg.rows != inputH || srcImg.cols != inputW) {
if (isClassification) {
cv::cuda::resize(imgRGB, resized, cv::Size(inputW, inputH), 0, 0, cv::INTER_LINEAR, stream);
cv::resize(srcImg, cpuResized, cv::Size(inputW, inputH), 0, 0, cv::INTER_LINEAR);
} else {
resized = Engine<float>::resizeKeepAspectRatioPadRightBottom(imgRGB, inputH, inputW);
cpuResized = Engine<float>::cpuResizeKeepAspectRatioPadRightBottom(srcImg, inputH, inputW);
}
}
else {
resized = imgRGB;
} else {
cpuResized = srcImg;
}
batchProcessed.push_back(std::move(resized));
cv::Mat cpuRGB;
cv::cvtColor(cpuResized, cpuRGB, cv::COLOR_BGR2RGB);
cv::cuda::GpuMat gpuResized;
gpuResized.upload(cpuRGB, stream);
batchProcessed.push_back(std::move(gpuResized));
}
stream.waitForCompletion();
@@ -1804,10 +1796,10 @@ namespace ANSCENTER {
std::vector<Object> ANSRTYOLO::DetectObjects(const cv::Mat& inputImage,
const std::string& camera_id) {
try {
// --- Debug timer helper (zero-cost when _debugFlag == false) ---
// --- Debug timer helper ---
using Clock = std::chrono::steady_clock;
const bool dbg = _debugFlag;
auto t0 = dbg ? Clock::now() : Clock::time_point{};
auto t0 = Clock::now(); // Always set — used by ANS_DBG timing output
auto tPrev = t0;
auto elapsed = [&]() -> double {
auto now = Clock::now();
@@ -2045,13 +2037,21 @@ namespace ANSCENTER {
}
// --- 6. Total pipeline time ---
if (dbg) {
{
double msTotal = std::chrono::duration<double, std::milli>(Clock::now() - t0).count();
_logger.LogInfo("ANSRTYOLO::DetectObjects",
"[DEBUG] " + camera_id + " | TOTAL=" + std::to_string(msTotal) +
"ms (" + std::to_string(inputImage.cols) + "x" + std::to_string(inputImage.rows) +
") Results=" + std::to_string(results.size()),
__FILE__, __LINE__);
if (dbg) {
_logger.LogInfo("ANSRTYOLO::DetectObjects",
"[DEBUG] " + camera_id + " | TOTAL=" + std::to_string(msTotal) +
"ms (" + std::to_string(inputImage.cols) + "x" + std::to_string(inputImage.rows) +
") Results=" + std::to_string(results.size()),
__FILE__, __LINE__);
}
// DebugView output — controlled by ANSCORE_DEBUGVIEW
double msPreproc = std::chrono::duration<double, std::milli>(_trtStart - t0).count();
ANS_DBG("YOLO_Timing", "cam=%s total=%.1fms preproc=%.1fms inf=%.1fms %dx%d det=%zu %s",
camera_id.c_str(), msTotal, msPreproc, _trtMs,
inputImage.cols, inputImage.rows, results.size(),
usedNV12 ? "NV12" : "BGR");
}
return results;
@@ -2101,7 +2101,7 @@ namespace ANSCENTER {
// --- Debug timer helper ---
using Clock = std::chrono::steady_clock;
const bool dbg = _debugFlag;
auto t0 = dbg ? Clock::now() : Clock::time_point{};
auto t0 = Clock::now(); // Always set — used by ANS_DBG timing output
auto tPrev = t0;
auto elapsed = [&]() -> double {
auto now = Clock::now();
@@ -2350,19 +2350,23 @@ namespace ANSCENTER {
}
}
if (dbg) {
double msPostprocess = elapsed();
{
double msPostprocess = dbg ? elapsed() : 0;
double msTotal = std::chrono::duration<double, std::milli>(Clock::now() - t0).count();
_logger.LogInfo("ANSRTYOLO::DetectObjectsBatch",
"[DEBUG] " + camera_id +
" batch=" + std::to_string(realCount) +
" | SetDev=" + std::to_string(msSetDevice) +
"ms Pad=" + std::to_string(msPad) +
"ms Preproc=" + std::to_string(msPreprocess) +
"ms Inf=" + std::to_string(msInference) +
"ms Postproc=" + std::to_string(msPostprocess) +
"ms TOTAL=" + std::to_string(msTotal) + "ms",
__FILE__, __LINE__);
if (dbg) {
_logger.LogInfo("ANSRTYOLO::DetectObjectsBatch",
"[DEBUG] " + camera_id +
" batch=" + std::to_string(realCount) +
" | SetDev=" + std::to_string(msSetDevice) +
"ms Pad=" + std::to_string(msPad) +
"ms Preproc=" + std::to_string(msPreprocess) +
"ms Inf=" + std::to_string(msInference) +
"ms Postproc=" + std::to_string(msPostprocess) +
"ms TOTAL=" + std::to_string(msTotal) + "ms",
__FILE__, __LINE__);
}
ANS_DBG("YOLO_Timing", "cam=%s batch=%d total=%.1fms preproc=%.1fms inf=%.1fms",
camera_id.c_str(), realCount, msTotal, msPreprocess, msInference);
}
return batchDetections;