Changelog

Markdown

Release notes summarized from the upstream tags and release commits. The full history is on GitHub; tagged releases are listed under tags.

Since v0.9.1 (unreleased)

Initial Cell Broadband Engine SPU back-end (SHZ_BACKEND=SHZ_SPU), vector routines only, added to CMake and to SHZ_TLS_MODEL auto-detection.
shz_vec_inv_sqrtf() and shz_vec_sqrtf() added to the type-generic routines and the C++ wrappers.
XMTRX, shz_fmodf() and shz_memcpyN() bug fixes; fixes to load_apply_store_4x4 and store_aligned32_4x4.
Build fixes for the KOS + SH toolchain stable versions and for Clang.

v0.9.1 (2026-09-25): XMTRX gainz, bug fixes, SW back-end on SH4

CMake: a configuration option chooses the back-end rather than deciding by architecture. The software back-end is now built for all targets, so a Dreamcast project can switch between the accelerated and software back-ends without rebuilding libsh4zam.a.
Makefile: SHZ_BACKEND variable to pick the SW or SH4 back-end.
XMTRX: faster SH4 versions of load_transpose_3x3, load_2x2, store_3x3, store_transpose_3x3, init_one, init_rotation_x/y/z, apply_translation, apply_scale, apply_rotation_x and init_screen. Software back-end fixes in apply_3x3, apply_transpose_3x3, apply_rotation_x/y/z, apply_rotation, apply_rotation_quat and rotate. New C++ wrappers for loading, storing and applying 2x2 and 3x3 matrices.
Memory: shz_sq_memcpy32() and shz_sq_memcpy32_1() software versions now work on real Dreamcast store queues.
cdefs: SHZ_TARGET is decoupled from SHZ_BACKEND, since any target can use either back-end; every major architecture is auto-detected as a target. -m4 and -m4-double-only SH4 builds are supported (with the software back-end).
Tests: benchmark gainz reported for the SW back-end too; coverage for every newly wrapped C++ routine and every routine that had a bug.

v0.9.0 (2026-09-22): XMTRX 3x4 API expansion

XMTRX 3x4 API for C and C++ on both back-ends: shz_xmtrx_load_transpose_3x4(), shz_xmtrx_store_transpose_3x4(), shz_xmtrx_apply_transpose_3x4(), shz_xmtrx_apply_reverse_3x4(), shz_xmtrx_apply_reverse_transpose_3x4() and the fused shz_xmtrx_load_apply_3x4().
The C++ API covers the whole 3x4 family.
Faster SH4 shz_xmtrx_get_scale(), shz_mat4x4_get_scale() (also fixing a real bug for non-affine matrices) and shz_mat4x4_transform_vec3(), which is now hand-written FIPR assembly.
More stable benchmark harness.

v0.8.1 (2026-09-20)

One-off rotations of a single 3D vector about the X, Y and Z axes: shz_vec3_rotate_x(), shz_vec3_rotate_y(), shz_vec3_rotate_z() (and shz::vec3::rotate_x() etc.), faster than going through XMTRX for a single rotation.
Resources page additions, including the deep-dive into FSCA, FSRRA, FTRV and FIPR.

v0.8.0 (2026-08-18)

"We're WAY overdue for a minor release." Lots of new features, many bug fixes for both back-ends, and no known regressions. Upstream recommends it as a stable tag to stay on before the 1.0 API.

v0.7.0 (2026-05-24)

Fix for software back-end unit test builds.