deepinsight/insightface · error · std::runtime_error

RGA buffer allocation failed

Error message

RGA buffer allocation failed

What it means

In the RGA image processor's buffer cache, when a GetBuffer cache miss occurs a new RGA buffer is allocated via buffer.Allocate(width, height, channels); if the underlying allocation (RGA/malloc/DMA memory) fails, the code throws std::runtime_error("RGA buffer allocation failed"). This aborts the copy/conversion step of the RGA pipeline. It almost always indicates resource exhaustion or dimensions the allocator rejects, not a logic bug in caller code.

Source

Thrown at cpp-package/inspireface/cpp/inspireface/image_process/nexus_processor/image_processor_rga.h:191

                last_dst_key_ = key;
            }
            return it->second;
        }

        if (buffer_cache_.size() >= 3) {  // Keep max 3 buffers in cache
            for (auto it = buffer_cache_.begin(); it != buffer_cache_.end();) {
                if (!(it->first == last_src_key_) && !(it->first == last_dst_key_)) {
                    it = buffer_cache_.erase(it);
                } else {
                    ++it;
                }
            }
        }

        auto& buffer = buffer_cache_[key];
        if (!buffer.Allocate(key.width, key.height, key.channels)) {
            INSPIRECV_LOG(ERROR) << "Failed to allocate RGA buffer";
            throw std::runtime_error("RGA buffer allocation failed");
        }

        if (is_src) {
            last_src_key_ = key;
        } else {
            last_dst_key_ = key;
        }

        return buffer;
    }

private:
    std::unordered_map<BufferKey, RGABuffer, BufferKeyHash> buffer_cache_;
    BufferKey last_src_key_{0, 0, 0};
    BufferKey last_dst_key_{0, 0, 0};
    int32_t aligned_width_{0};
};

View on GitHub (pinned to 7fadd420c2)

Solutions

1. Free memory / reduce concurrent load: close other camera or inference processes and retry to see if allocation succeeds.
2. Reduce input resolution or reuse a consistent frame size so the buffer cache hits (buffer_cache_[key]) instead of allocating new buffers every frame.
3. Check RGA availability and memory: verify /dev/rga exists, the rga driver is loaded, and inspect CMA/DMA heap usage (dmesg | grep -i rga, /proc/meminfo).
4. If the leak is repeated cache growth, bound or clear the buffer cache between sessions and report upstream if buffers are never released.

Example fix

// before
auto& proc = ...; // ImageProcessorRGA
proc->CopyImage(img, dst); // may throw std::runtime_error: RGA buffer allocation failed
// after
try {
    proc->CopyImage(img, dst);
} catch (const std::runtime_error& e) {
    LOG(WARNING) << "RGA allocation failed: " << e.what()
                 << " — retrying with downscaled frame";
    auto small = img.Resize(img.Width()/2, img.Height()/2);
    proc->CopyImage(small, dst);
}

size_t need = (size_t)width * height * channels;
if (need > kMaxRgaBufferBytes) { /* downscale or reject frame before processing */ }

try {
    processor->CopyImage(src, dst);
} catch (const std::runtime_error& e) {
    if (std::string(e.what()).find("RGA buffer") != std::string::npos) {
        // free memory / downscale input and retry once
    } else throw;
}

When it happens

Trigger: Requesting a conversion/copy with unusually large width×height×channels so the RGA buffer allocation exceeds available memory; many concurrent streams creating distinct cache keys and fragmenting/exhausting the cache; running on a device where the RGA device (/dev/rga) or its DMA heap is unavailable or out of memory; memory pressure on embedded boards after long uptime or alongside other camera/AI processes.

Common situations: Feeding 4K/high-resolution frames on RK3588/RK3566 boards where CMA or DMA-BUF heap is exhausted; multiple cameras or pipeline instances each allocating new buffer keys because width/height/format differ per frame; OOM conditions on Android/Embedded Linux when InspireFace runs next to other heavy media processes; a first-run failure because the RGA driver isn't loaded.

AI-assisted analysis of deepinsight/insightface@7fadd420c2 (2026-08-28). Data as JSON: /api/errors/c5b2ff9d30ea4ac5. Report an issue: GitHub.