v0.2.17

Fixed

• Evision.Mat.clip/3: upper bound no longer collapses to the lower bound (typo where both branches compared against lower).
• Evision.Mat.dot/2: now actually computes a dot product. The previous implementation called cv::Mat::cross despite being named dot, so it only worked for 3-element vectors and returned a Mat. See Breaking changes below.
• Evision.Mat.arange/4,5: range count is now computed with ceiling division, so the final stepped value is included when the interval is not evenly divisible (e.g. arange(0, 5, 2) now returns [0, 2, 4] instead of [0, 2]). step == 0 and direction mismatches are rejected explicitly.
• Evision.Mat.negate/1, Evision.Mat.ceil/1, Evision.Mat.floor/1, Evision.Mat.round/1: now traverse every scalar in multi-channel matrices instead of only the first channel of each pixel.
• Evision.Mat.ceil/1, Evision.Mat.floor/1, Evision.Mat.round/1: added support for CV_16F (half-precision float) inputs.
• Evision.Mat.at/2: added bounds checking (was previously undefined behaviour past the end) and a clone-on-non-continuous fallback so the scalar index lands on the right element. CV_16F is now supported.
• Evision.Mat.from_binary_by_shape/3: rejects binaries whose byte size does not match the declared shape (previously created a Mat that read past the binary).
• Evision.Mat.from_binary/5: the declared rows × cols × channels × bytes_per_element is now computed with overflow checks.
• Evision.Mat.to_binary/1,2:
  • The limit parameter is now honored on the NIF path (was previously ignored).
  • Non-continuous matrices are cloned before serialisation so the returned binary contains contiguous element data rather than the underlying buffer's interleaved padding.
  • The nx_tensor branch in Elixir now interprets limit as the number of elements (multiplying by elem_size), matching the NIF path and the Backend.inspect/2 caller.
• Evision.Mat.to_batched/3 (3-arg): now works on multi-channel matrices (previously the shape(mat) it derived internally always failed the size check). The 3-arg wrapper also no longer passes a tuple where a list is expected.
• Evision.Mat.to_batched/4 repeat mode: no longer reads past the source matrix when batch_size × slice_size exceeds the source size; the source is now safely repeated as many times as needed.
• Evision.Mat.update_roi/3: the caller's Mat data is no longer mutated in place — update_roi always returns a new Mat and leaves the input untouched. Previously the underlying buffer was modified through the shared reference, silently mutating any other reference to the same Mat.
• Evision.Mat.transpose/3 with as_shape:: rejects as_shape whose total byte size does not match the source matrix, preventing out-of-bounds reads from cv::transposeND.
• Evision.Mat.expm1/1: removed a dead/buggy fallback that called cv::exp a second time without exception protection after the wrapped call had already failed.
• Evision.Mat.last_dim_as_channel/1: now uses src.depth() (instead of the full type) when re-tagging the channels and explicitly rejects channel counts outside [1, CV_CN_MAX].
• Evision.CUDA.GpuMat.from_pointer/3: the resource is now released on the error path (previously leaked) and the OOM path returns a proper error term.
• evision::nif::get_tuple for std::vector<int64_t>: now reads each element as int64_t (was int), so large 64-bit values in tuples are no longer truncated/rejected.
• evision_highgui thread initialisation: races between concurrent start_native_gui callers are now serialised behind an EVISION_INITIATING state, and evision_main_loop notifies all waiters (was notify_one, which could leave additional waiters stuck).
• alloc_resource for cv::Mat *: explicitly initialises tmp->val = nullptr so a release before assignment does not free uninitialised memory.

Added

• Evision.Mat.matmul/2: convenience wrapper around cv::Mat::operator* (matrix multiplication via cv::gemm). Both inputs must be 2-D and share a CV_32F or CV_64F depth. For element-wise dot-product semantics use Evision.Mat.dot/2.

Breaking changes

• Evision.Mat.dot/2 now returns a number() (the scalar dot product) instead of a %Evision.Mat{}. This matches the C++ cv::Mat::dot definition. The previous behaviour was a bug: the function was implemented with cv::Mat::cross, which only accepted 3-element vectors and returned a Mat. Callers who actually wanted matrix multiplication should switch to the new Evision.Mat.matmul/2.
• Evision.Mat.to_batched/4 (4-arg, explicit as_shape) now requires as_shape to include channels as a dimension for multi-channel matrices, and the resulting batches are single-channel Mats with the channel axis at the end. Concretely:
  • Old: to_batched(cv_8uc3_mat, batch_size, {N, H, W}, ...) → batches of CV_8UC3 mats with shape {batch_size, H, W}.
  • New: to_batched(cv_8uc3_mat, batch_size, {N, H, W, 3}, ...) → batches of CV_8U mats with shape {batch_size, H, W, 3}.
  • The old to_batched(cv_8uc3_mat, batch_size, {N, H, W}, ...) call now returns {:error, "to_batched failed: cannot treated matrix as the request shape"}.
  • Evision.Backend (Nx) workflows are not affected: tensors entering through from_nx/from_binary are always single-channel Mats with the logical Nx shape, so the old and new size-checks evaluate identically.