Fix multi-GPU inference on Blackwell (dual RTX 5080): four stacked fixes
Verified on dual RTX 5080: 355/355 detections, both GPUs in round-robin, zero CUDA errors. Requires TensorRT >= 10.16.1 in SharedCore (10.15.1 has a Myelin l2cm crash when creating a context on a second Blackwell device). 1. blobFromGpuMats: thread_local GpuMat caches (tl_blob/tl_floatImg) were device-blind — a pool thread warming GPU[1] then GPU[0] reused device-1 memory with a device-0 stream, causing an async illegal memory access that poisoned every CUDA context (sticky 700). Track owning device and rebuild the cache on device switch. 2. runInference: pool round-robin can dispatch to a slot on a different GPU than the one the caller prepared inputs on. Detect foreign input GpuMats via cudaPointerGetAttributes and migrate them through a pinned host staging buffer (two-hop D2H+H2D; direct cross-device cudaMemcpy2D is unsupported and GeForce has no P2P). 3. warmUp: now pins its own CUDA device before uploading dummy inputs and returns success/failure; loadNetwork rejects engines whose warmup fails so broken slots can never join the pool silently. 4. Blackwell builder workaround: setMaxAuxStreams(0) on SM 12.x plus an .aux0 engine-cache suffix so pre-workaround caches rebuild. Also: ANSFR-UnitTest gains --gpu N flag (sets CUDA_VISIBLE_DEVICES) for single-GPU bisection, and rich diagnostics on input-copy failures. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
@@ -130,13 +130,31 @@ static inline cudaError_t cudaStreamSynchronize_Safe(cudaStream_t stream) {
|
||||
}
|
||||
|
||||
template <typename T>
|
||||
void Engine<T>::warmUp(int iterations) {
|
||||
bool Engine<T>::warmUp(int iterations) {
|
||||
if (m_verbose) {
|
||||
std::cout << "\n========================================" << std::endl;
|
||||
std::cout << "Engine Warmup" << std::endl;
|
||||
std::cout << "========================================" << std::endl;
|
||||
}
|
||||
|
||||
// Ensure this thread's CUDA device matches the engine BEFORE the dummy-input
|
||||
// uploads below. warmUp() does GPU work (GpuMat uploads) ahead of
|
||||
// runInference()'s own device guard, so it must not rely on the caller
|
||||
// having left the right device current (multi-GPU pool creation touches
|
||||
// several devices on one thread).
|
||||
{
|
||||
int currentDev = -1;
|
||||
cudaGetDevice(¤tDev);
|
||||
if (currentDev != m_options.deviceIndex) {
|
||||
std::cout << "Warning [Warmup]: thread was on GPU[" << currentDev
|
||||
<< "], engine expects GPU[" << m_options.deviceIndex
|
||||
<< "] -- switching device" << std::endl;
|
||||
cudaSetDevice(m_options.deviceIndex);
|
||||
}
|
||||
}
|
||||
|
||||
bool allBatchesOk = true;
|
||||
|
||||
// Determine batch sizes to warm up
|
||||
std::vector<int> batchSizes;
|
||||
|
||||
@@ -212,7 +230,9 @@ void Engine<T>::warmUp(int iterations) {
|
||||
auto firstTime = std::chrono::duration_cast<std::chrono::milliseconds>(end - start).count();
|
||||
|
||||
if (!firstSuccess) {
|
||||
if (m_verbose) std::cout << " ✗ ERROR: First iteration failed for batch=" << batchSize << std::endl;
|
||||
std::cout << " ✗ ERROR: Warmup first iteration failed for batch=" << batchSize
|
||||
<< " on GPU[" << m_options.deviceIndex << "]" << std::endl;
|
||||
allBatchesOk = false;
|
||||
continue;
|
||||
}
|
||||
|
||||
@@ -231,7 +251,9 @@ void Engine<T>::warmUp(int iterations) {
|
||||
for (int i = 1; i < iters; ++i) {
|
||||
bool success = runInference(dummyInputs, dummyOutputs);
|
||||
if (!success) {
|
||||
if (m_verbose) std::cout << " ✗ ERROR: Iteration " << i << " failed" << std::endl;
|
||||
std::cout << " ✗ ERROR: Warmup iteration " << i << " failed for batch="
|
||||
<< batchSize << " on GPU[" << m_options.deviceIndex << "]" << std::endl;
|
||||
allBatchesOk = false;
|
||||
break;
|
||||
}
|
||||
}
|
||||
@@ -262,6 +284,7 @@ void Engine<T>::warmUp(int iterations) {
|
||||
std::cout << "Kernels pre-compiled for all batch sizes." << std::endl;
|
||||
std::cout << "========================================\n" << std::endl;
|
||||
}
|
||||
return allBatchesOk;
|
||||
}
|
||||
template <typename T>
|
||||
bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& inputs,std::vector<std::vector<std::vector<T>>>& featureVectors) {
|
||||
@@ -778,6 +801,118 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
|
||||
}
|
||||
}
|
||||
|
||||
// Pass 1.5: INPUT DEVICE MIGRATION (multi-GPU pool dispatch).
|
||||
//
|
||||
// The caller prepares input GpuMats on ITS device (e.g. camera preprocessing
|
||||
// on GPU 0), but round-robin dispatch may select a slot on another GPU.
|
||||
// Kernels cannot touch cross-device memory without peer access: preprocessing
|
||||
// a device-A GpuMat with a device-B stream raises an async illegal memory
|
||||
// access that permanently poisons every CUDA context in the process
|
||||
// (sticky error 700). Detect foreign inputs and migrate them onto this
|
||||
// engine's device via a UVA copy (cudaMemcpyDefault routes over PCIe).
|
||||
std::vector<std::vector<cv::cuda::GpuMat>> migratedInputs;
|
||||
const std::vector<std::vector<cv::cuda::GpuMat>>* activeInputs = &inputs;
|
||||
{
|
||||
bool anyForeign = false;
|
||||
for (size_t i = 0; i < numInputs && !anyForeign; ++i) {
|
||||
for (const auto& img : inputs[i]) {
|
||||
if (!img.data) continue;
|
||||
cudaPointerAttributes attr{};
|
||||
if (cudaPointerGetAttributes(&attr, img.data) == cudaSuccess &&
|
||||
attr.type == cudaMemoryTypeDevice &&
|
||||
attr.device != m_options.deviceIndex) {
|
||||
anyForeign = true;
|
||||
}
|
||||
cudaGetLastError(); // clear any error the attribute query left behind
|
||||
break; // first image per tensor is representative
|
||||
}
|
||||
}
|
||||
|
||||
if (anyForeign) {
|
||||
migratedInputs.resize(numInputs);
|
||||
for (size_t i = 0; i < numInputs; ++i) {
|
||||
migratedInputs[i].reserve(inputs[i].size());
|
||||
for (const auto& src : inputs[i]) {
|
||||
cudaPointerAttributes attr{};
|
||||
const bool foreign = src.data &&
|
||||
cudaPointerGetAttributes(&attr, src.data) == cudaSuccess &&
|
||||
attr.type == cudaMemoryTypeDevice &&
|
||||
attr.device != m_options.deviceIndex;
|
||||
cudaGetLastError();
|
||||
if (!foreign) { migratedInputs[i].push_back(src); continue; }
|
||||
|
||||
// Two-hop copy through pinned host staging. A direct
|
||||
// device-to-device cudaMemcpy2D across GPUs is not supported
|
||||
// (2D copies have no peer path, and GeForce has no P2P), so:
|
||||
// hop 1: strided D2H from the foreign GPU into pinned host
|
||||
// hop 2: strided H2D onto this engine's GPU
|
||||
// Synchronous copies keep the single staging buffer safe to
|
||||
// reuse across images; ~1 ms per 1080p frame over PCIe 4.
|
||||
struct PinnedStaging { // thread-local, grow-only
|
||||
void* ptr = nullptr;
|
||||
size_t cap = 0;
|
||||
~PinnedStaging() { if (ptr) cudaFreeHost(ptr); }
|
||||
};
|
||||
thread_local PinnedStaging tl_staging;
|
||||
|
||||
const size_t rowBytes = static_cast<size_t>(src.cols) * src.elemSize();
|
||||
const size_t needed = rowBytes * static_cast<size_t>(src.rows);
|
||||
if (needed > tl_staging.cap) {
|
||||
if (tl_staging.ptr) cudaFreeHost(tl_staging.ptr);
|
||||
tl_staging.ptr = nullptr; tl_staging.cap = 0;
|
||||
if (cudaMallocHost(&tl_staging.ptr, needed) != cudaSuccess) {
|
||||
std::cout << "Error [MultiGPU]: pinned staging alloc failed ("
|
||||
<< needed << " bytes)" << std::endl;
|
||||
return false;
|
||||
}
|
||||
tl_staging.cap = needed;
|
||||
}
|
||||
|
||||
// Hop 1: D2H. UVA lets us read the foreign pointer from this
|
||||
// device's context; if the runtime rejects it, retry with the
|
||||
// source device current.
|
||||
cudaError_t mErr = cudaMemcpy2D(
|
||||
tl_staging.ptr, rowBytes, src.data, src.step,
|
||||
rowBytes, src.rows, cudaMemcpyDeviceToHost);
|
||||
if (mErr != cudaSuccess) {
|
||||
cudaGetLastError();
|
||||
cudaSetDevice(attr.device);
|
||||
mErr = cudaMemcpy2D(
|
||||
tl_staging.ptr, rowBytes, src.data, src.step,
|
||||
rowBytes, src.rows, cudaMemcpyDeviceToHost);
|
||||
cudaSetDevice(m_options.deviceIndex);
|
||||
}
|
||||
if (mErr != cudaSuccess) {
|
||||
std::cout << "Error [MultiGPU]: migration hop1 (D2H from GPU["
|
||||
<< attr.device << "]) failed: " << cudaGetErrorString(mErr) << std::endl;
|
||||
return false;
|
||||
}
|
||||
|
||||
// Hop 2: H2D onto this engine's GPU.
|
||||
cv::cuda::GpuMat dst(src.rows, src.cols, src.type());
|
||||
mErr = cudaMemcpy2D(
|
||||
dst.data, dst.step, tl_staging.ptr, rowBytes,
|
||||
rowBytes, src.rows, cudaMemcpyHostToDevice);
|
||||
if (mErr != cudaSuccess) {
|
||||
std::cout << "Error [MultiGPU]: migration hop2 (H2D to GPU["
|
||||
<< m_options.deviceIndex << "]) failed: "
|
||||
<< cudaGetErrorString(mErr) << std::endl;
|
||||
return false;
|
||||
}
|
||||
migratedInputs[i].push_back(dst);
|
||||
}
|
||||
}
|
||||
activeInputs = &migratedInputs;
|
||||
|
||||
static std::atomic<bool> s_migrationLogged{false};
|
||||
if (!s_migrationLogged.exchange(true)) {
|
||||
std::cout << "Info [MultiGPU]: inputs resided on another GPU -- migrated to GPU["
|
||||
<< m_options.deviceIndex << "] via UVA copy (logged once; recurs per"
|
||||
<< " cross-device dispatch)" << std::endl;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
// Pass 2: Preprocess + D2D copies — all on m_inferenceStream (no null stream).
|
||||
//
|
||||
// All OpenCV CUDA ops (convertTo, subtract, divide, split) in blobFromGpuMats
|
||||
@@ -794,7 +929,7 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
|
||||
preprocessedBuffers.reserve(numInputs);
|
||||
|
||||
for (size_t i = 0; i < numInputs; ++i) {
|
||||
const auto& batchInput = inputs[i];
|
||||
const auto& batchInput = (*activeInputs)[i];
|
||||
|
||||
// Preprocess on m_inferenceStream (not the null stream).
|
||||
preprocessedBuffers.push_back(
|
||||
@@ -811,8 +946,23 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
|
||||
m_inferenceStream);
|
||||
|
||||
if (copyErr != cudaSuccess) {
|
||||
// Rich diagnostics — this fires on multi-GPU slot warmup failures
|
||||
// (Blackwell debugging). Identify WHICH precondition broke:
|
||||
// wrong current device, empty blob (failed allocation), null dest
|
||||
// buffer, or a pre-existing sticky async error.
|
||||
int curDev = -1; cudaGetDevice(&curDev);
|
||||
std::cout << "Error: Failed to copy input " << i
|
||||
<< " to inference buffer: " << cudaGetErrorString(copyErr) << std::endl;
|
||||
<< " to inference buffer: " << cudaGetErrorString(copyErr) << "\n"
|
||||
<< " [diag] engine deviceIndex = " << m_options.deviceIndex
|
||||
<< ", thread current device = " << curDev << "\n"
|
||||
<< " [diag] src blob ptr = " << static_cast<const void*>(blobMat.data)
|
||||
<< " (" << blobMat.rows << "x" << blobMat.cols
|
||||
<< ", elemSize " << blobMat.elemSize() << ")\n"
|
||||
<< " [diag] dst buffer ptr = " << m_buffers[i]
|
||||
<< ", copySize = " << copySize << " bytes\n"
|
||||
<< " [diag] inference stream = " << static_cast<const void*>(m_inferenceStream)
|
||||
<< ", pending lastError = " << cudaGetErrorString(cudaGetLastError())
|
||||
<< std::endl;
|
||||
return false;
|
||||
}
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user