-1 = error

! * @internal * @brief Processes a large input for XXH3_update() and XXH3_digest_long(). * * Unlike XXH3_hashLong_internal_loop(), this can proces

! * @brief Calculates a 64-bit to 128-bit multiply, then XOR folds it. * * The reason for the separate function is to prevent passing too many stru

! @ingroup XXH64_family */

! * Performs an unaligned vector load and byte swaps it on big endian. */

Allocate buffer for literals, either overlapping current dst, or split between dst and litExtraBuffer, or stored entirely within litExtraBuffer */

! ZSTD_buildSeqTable() : * @return : nb bytes read from src, * or an error code if it fails */

Need to use memmove here since the literal buffer can now be located within the dst buffer. In circumstances where the op "catches up" to where the

ZSTD_decodeFrameHeader() : * `headerSize` must be the size provided by ZSTD_frameHeaderSize(). * If multiple DDict references are enabled, also will

ZSTD_getFrameContentSize() : * compatible with legacy mode * @return : decompressed size of the single frame pointed to be `src` if known, otherwis

! ZSTD_getcBlockSize() : * Provides the size of compressed block from block header `src` */

* Similar to ZSTD_nextSrcSizeToDecompress(), but when a block input can be streamed, we * allow taking a partial block as the input. Currently only r

! ZSTD_overlapCopy8() : * Copies 8 bytes from ip to op and updates op and ip where ip <= op. * If the offset is < 8 then the offset is spread to a

! ZSTD_safecopy() : * Specialized version of memcpy() that is allowed to READ up to WILDCOPY_OVERLENGTH past the input buffer * and write up to 16

Add a custom error message

MODE 0 0. DualPlane: 0, WeightRange: 8 (16), Subsets: 1, CEM: 8 (RGB Direct ), EndpointRange: 19 (192) MODE6 RGB 18. DualPlane: 0, WeightR

input: a=[0,255] b=[0,255] output: a=from, converted to -32 to 31 b=to, shifted right by 1 and 1 bit added to MSB, so [0,255]

TODO: Try two-step Lanczos iteration/Rayleigh�Ritz approximation in a 2-dimensional Krylov subspace method vs. power method.

May fail if the tiebreaking logic isn't strong enough.

clear() sets the container to empty, then frees the allocated block.

Reference: https://ece.uwaterloo.ca/~z70wang/research/ssim/index.html

endpoints is [4][2]

Assumes the decode_unorm8 extension is active (only upper 8 bits used).

This float->half conversion matches how "F32TO16" works on Intel GPU's.