fix(ffs-depth-tensorrt): depth_alloc_pixels_ not reset (+2 more)

[andrewwhitecdw]

Small fixes in cpp/src/ffs_depth_tensorrt.cpp:

fix: depth_alloc_pixels_ not reset before realloc

Fix: Replace:

    if (static_cast<int64_t>(num_pixels) > depth_alloc_pixels_) {
        // cudaFree is implicitly stream-ordered: it waits for prior work on the
        // device before reclaiming the allocation, so this is safe to call here
        // even though earlier inferDepth() calls may still be in flight.
        if (d_disp_for_depth_) {
            cudaFree(d_disp_for_depth_);
            d_disp_for_depth_ = nullptr;
        }
        cudaMallocChecked(reinterpret_cast<void**>(&d_disp_for_depth_),
                          num_pixels * sizeof(float), "d_disp_for_depth_");
        depth_alloc_pixels_ = static_cast<int64_t>(num_pixels);
    }

with:

    if (static_cast<int64_t>(num_pixels) > depth_alloc_pixels_) {
        // cudaFree is implicitly stream-ordered: it waits for prior work on the
        // device before reclaiming the allocation, so this is safe to call here
        // even though earlier inferDepth() calls may still be in flight.
        if (d_disp_for_depth_) {
            cudaFree(d_disp_for_depth_);
            d_disp_for_depth_ = nullptr;
        }
        depth_alloc_pixels_ = 0;
        cudaMallocChecked(reinterpret_cast<void**>(&d_disp_for_depth_),
                          num_pixels * sizeof(float), "d_disp_for_depth_");
        depth_alloc_pixels_ = static_cast<int64_t>(num_pixels);
    }

fix: inferDepth does not validate output pointer

Fix: Replace:

void FFSDepthInference::inferDepth(
    const uint8_t* d_left_rgb, const uint8_t* d_right_rgb,
    int input_h, int input_w,
    float fx, float baseline_m,
    float* d_depth_out)
{
    if (input_h <= 0 || input_w <= 0) {
        throw std::runtime_error("[FFS] inferDepth: input dimensions must be positive");
    }

with:

void FFSDepthInference::inferDepth(
    const uint8_t* d_left_rgb, const uint8_t* d_right_rgb,
    int input_h, int input_w,
    float fx, float baseline_m,
    float* d_depth_out)
{
    if (!d_left_rgb || !d_right_rgb || !d_depth_out) {
        throw std::runtime_error("[FFS] inferDepth: null device pointer");
    }
    if (input_h <= 0 || input_w <= 0) {
        throw std::runtime_error("[FFS] inferDepth: input dimensions must be positive");
    }

fix: disp buffer logic assumes FP32 regardless of dtype

Fix: Replace:

    const auto disp_dt = post_engine_->getTensorDataType("disp");
    const size_t disp_bytes = H * W * elementSize(disp_dt);
    cudaMallocChecked(reinterpret_cast<void**>(&d_disp_), disp_bytes, "d_disp_");

with:

    const auto disp_dt = post_engine_->getTensorDataType("disp");
    if (disp_dt != nvinfer1::DataType::kFLOAT) {
        throw std::runtime_error("[FFS] post_runner disp tensor must be FP32");
    }
    const size_t disp_bytes = H * W * elementSize(disp_dt);
    cudaMallocChecked(reinterpret_cast<void**>(&d_disp_), disp_bytes, "d_disp_");

Files changed

• cpp/src/ffs_depth_tensorrt.cpp