Release/0.5.2

.github/workflows/release.yml

@@ -0,0 +1,101 @@
+# Prebuilt binaries, so getting Booth does not require a C toolchain and a
+# working `make`. Runs on a version tag, or by hand from the Actions tab if
+# you want to check the packaging without tagging anything.
+name: Release binaries
+
+on:
+  push:
+    tags: ['v*']
+  workflow_dispatch:
+
+permissions:
+  contents: write
+
+jobs:
+  build:
+    name: ${{ matrix.name }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          # -static on Linux and Windows gives a binary with no runtime
+          # dependencies at all. Without it the Windows build wants MinGW's
+          # libssp-0.dll, which nobody downloading a release will have.
+          - { os: ubuntu-latest,  name: linux-x86_64,   ld: '-static', ext: ''    }
+          - { os: windows-latest, name: windows-x86_64, ld: '-static', ext: '.exe' }
+          # Apple ships no static libc, so this one links against libSystem.
+          # That is fine, every macOS has it.
+          - { os: macos-latest,   name: macos-arm64,    ld: '',        ext: ''    }
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up MSYS2
+        if: runner.os == 'Windows'
+        uses: msys2/setup-msys2@v2
+        with:
+          msystem: MINGW64
+          install: mingw-w64-x86_64-gcc make
+
+      - name: Build
+        shell: ${{ runner.os == 'Windows' && 'msys2 {0}' || 'bash' }}
+        run: make LDFLAGS="${{ matrix.ld }}" -j4
+
+      - name: Smoke test
+        shell: ${{ runner.os == 'Windows' && 'msys2 {0}' || 'bash' }}
+        run: |
+          ./kath${{ matrix.ext }} --version
+          printf '__global__ void k(float *o){ o[0] = 1.0f; }\n' > smoke.cu
+          ./kath${{ matrix.ext }} --nvidia-ptx smoke.cu -o smoke.ptx
+          test -s smoke.ptx
+          echo "compiled a kernel, binary works"
+
+      # A release archive is the binary plus the message catalogues --lang
+      # reads and the licence. Everything else is source.
+      - name: Package
+        shell: ${{ runner.os == 'Windows' && 'msys2 {0}' || 'bash' }}
+        run: |
+          VER=$(awk '$2 == "BC_VERSION_STRING" {gsub(/"/, "", $3); print $3}' src/barracuda.h)
+          DIR="booth-$VER-${{ matrix.name }}"
+          mkdir -p "$DIR/lang"
+          cp kath${{ matrix.ext }} "$DIR/"
+          cp lang/en.txt lang/mi.txt "$DIR/lang/"
+          cp LICENSE README.md "$DIR/"
+          if [ "${{ runner.os }}" = "Windows" ]; then
+            powershell -Command "Compress-Archive -Path '$DIR' -DestinationPath '$DIR.zip'"
+          else
+            tar czf "$DIR.tar.gz" "$DIR"
+          fi
+          ls -la booth-$VER-*
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ${{ matrix.name }}
+          path: |
+            booth-*.tar.gz
+            booth-*.zip
+          if-no-files-found: error
+
+  publish:
+    name: Attach to release
+    needs: build
+    runs-on: ubuntu-latest
+    if: startsWith(github.ref, 'refs/tags/')
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          merge-multiple: true
+
+      # Checksums so anyone who cares can verify what they downloaded.
+      - name: Checksums
+        run: |
+          ls -la
+          sha256sum booth-* > SHA256SUMS
+          cat SHA256SUMS
+
+      - uses: softprops/action-gh-release@v2
+        with:
+          files: |
+            booth-*
+            SHA256SUMS

CHANGELOG.md

@@ -3,6 +3,67 @@ Booth — Changelog
 
 ## Unreleased
 
+## 2026-08-07
+
+Version 0.5.2.
+
+First, a correction. The last release went out tagged v5.01, which was
+meant to be 0.5.1 and wasn't, and it left Booth looking four major
+versions further along than it actually is. It isn't. This release puts
+the numbering back where it belongs, and sorry to anyone who pinned the
+old one or took the version at face value. The tag stays where it is so
+nothing breaks underneath you, but the compiler now reports what it is.
+
+The theme this cycle, without meaning to be, was the compiler telling
+the truth. Semantic errors used to be printed and then ignored by every
+mode except `--sema`, so the backend ran on source that had already been
+rejected, wrote an output file and exited zero. Asking for several
+backends at once wrote all of them over the same `-o` path and left you
+whichever finished last, under the name you chose, again exiting zero.
+Metal quietly narrowed a double-precision kernel to float and said
+nothing, which is a real problem if you were counting on the precision.
+`--amdgpu` ignored `-o` entirely and mixed a diagnostic into the
+assembly on stdout. All four are fixed, and all four had been sitting
+there being cheerfully wrong for a while.
+
+The structural change is the backend contract. Every target now sits
+behind a `be_desc_t` and registers itself in one list, and a backend
+owns its own command line rather than reaching into a shared config
+struct and the driver's argument loop. Adding a target used to mean
+reading 420 KB of AMD backend to work out what was expected; now it
+means reading one header and copying the skeleton. `main.c` lost about
+a third of its length in the process, and the frontend stopped needing
+to know what an AMD target enum is.
+
+Booth also installs now. `make install` puts `kath`, the message
+catalogues and a CMake package config into a prefix, so a downstream
+project can `find_package(Booth)` and compile kernels as part of its own
+build with `booth_add_kernel()`. There is a worked example under
+`examples/cmake/` and CI builds it against a staged install on every
+push, along with a check that the target list in the package config
+hasn't drifted from what the backends actually accept.
+
+Getting Booth no longer requires being able to build it. Every release
+now carries prebuilt binaries for Linux, macOS and Windows, statically
+linked so they have no runtime dependencies whatsoever, with checksums.
+Unpack and run. This mattered more than I realised: the Windows build
+had been quietly depending on MinGW's `libssp-0.dll`, so handing someone
+the binary would not have worked on a machine without a toolchain, which
+is exactly the machine they wanted it for.
+
+There are coverage numbers for the first time, 74.1% of lines and 58.4%
+of branches, reported by CI on every PR. That immediately turned up the
+SSA register allocator having never been executed by a test at all, and
+the six fixtures now pinning its behaviour also pin a real bug in it,
+which is at least honest.
+
+Thanks to Jorge Galvez, whose do-concurrent ocean benchmarks found three
+genuine frontend bugs in an afternoon, and who let me test against his
+code. Thanks to @maou3434 for the bare HIP warp and lane intrinsics,
+which is their first contribution here and a very welcome one. And
+thanks to @FileDelta for asking a simple question about the runtimes
+that turned over considerably more than either of us expected.
+
 ### Frontend
 
 - double-precision `fmax`, `fmin` and `fmod`. The ocean kernels in
@@ -145,6 +206,13 @@ Booth — Changelog
   kernels with `booth_add_kernel()`
   (Zane Hambly, 2026-08-06)
 
+- prebuilt binaries on every release for Linux, macOS and Windows, statically
+  linked where the platform allows it so they carry no runtime dependencies,
+  with SHA256 checksums. The Windows build previously needed MinGW's
+  `libssp-0.dll` present, which made the binary useless to anyone without a
+  toolchain
+  (Zane Hambly, 2026-08-07)
+
 ### CI and tests
 
 - #140: numeric regression against SLATEC known-good values, across cpu,
@@ -194,7 +262,8 @@ Booth — Changelog
 
 ## 2026-07-14
 
-Version 5.01. So long, and thanks for all the fish.
+Version 5.01, which should have read 0.5.1. See the 0.5.2 note above.
+So long, and thanks for all the fish.
 BarraCUDA was a good pun and a bad description, so it swam off: the
 compiler is Booth now, the binary is `kath`, and the version jumps
 to mark the line. First release under the new name; the rename note

Makefile

@@ -188,8 +188,10 @@ install: $(TARGET)
 	install -m 755 $(TARGET)$(EXE) $(BINDIR)/$(TARGET)$(EXE)
 	install -m 644 lang/en.txt lang/mi.txt $(SHAREDIR)/lang/
 	sed -e 's/@BOOTH_VERSION@/$(VERSION)/g' -e 's/@BOOTH_VERSION_MAJOR@/$(VER_MAJOR)/g' \
+	    -e 's/@BOOTH_VERSION_MINOR@/$(VER_MINOR)/g' \
 	    cmake/BoothConfig.cmake.in > $(CMAKEDIR)/BoothConfig.cmake
 	sed -e 's/@BOOTH_VERSION@/$(VERSION)/g' -e 's/@BOOTH_VERSION_MAJOR@/$(VER_MAJOR)/g' \
+	    -e 's/@BOOTH_VERSION_MINOR@/$(VER_MINOR)/g' \
 	    cmake/BoothConfigVersion.cmake.in > $(CMAKEDIR)/BoothConfigVersion.cmake
 
 uninstall:

README.md

@@ -18,8 +18,20 @@ That last one still surprises me a bit. You can write a Triton kernel, matmul an
 
 It also borrows a pile of operational discipline from the mainframe world: real crash dumps when a kernel faults, structured output routed by class, parameter snapshots on entry. See [docs/mainframe.md](docs/mainframe.md) if that sounds like your kind of thing.
 
+## Getting it
+
+If you're lazy like me and just want to run and go, then you can install it [here](https://github.com/Zaneham/Booth/releases/latest). It comes with no dependencies and you don't have to run `make` to use it. There's a build for Linux, macOS and Windows.
+
+```bash
+tar xzf booth-*-linux-x86_64.tar.gz
+cd booth-*-linux-x86_64
+./kath --version
+```
+
 ## Build
 
+If you'd rather build it, or you're on something I don't ship a binary for:
+
 ```bash
 make
 ```

cmake/BoothConfigVersion.cmake.in

@@ -1,18 +1,23 @@
-# Version check for find_package(Booth <version>). Same-major compatibility:
-# a consumer asking for 5.1 is satisfied by any later 5.x, never by 6.x.
+# Version check for find_package(Booth <version>).
+#
+# While the major version is 0 the minor has to match too, because 0.x is
+# where the shape is still allowed to change and 0.6 promises nothing to
+# someone who asked for 0.5. From 1.0 onward the usual same-major rule
+# applies and a consumer asking for 1.2 is happy with any later 1.x.
 
 set(PACKAGE_VERSION "@BOOTH_VERSION@")
 
 if(PACKAGE_VERSION VERSION_LESS PACKAGE_FIND_VERSION)
     set(PACKAGE_VERSION_COMPATIBLE FALSE)
+elseif(NOT PACKAGE_FIND_VERSION_MAJOR STREQUAL "@BOOTH_VERSION_MAJOR@")
+    set(PACKAGE_VERSION_COMPATIBLE FALSE)
+elseif("@BOOTH_VERSION_MAJOR@" STREQUAL "0" AND
+       NOT PACKAGE_FIND_VERSION_MINOR STREQUAL "@BOOTH_VERSION_MINOR@")
+    set(PACKAGE_VERSION_COMPATIBLE FALSE)
 else()
-    if(PACKAGE_FIND_VERSION_MAJOR STREQUAL "@BOOTH_VERSION_MAJOR@")
-        set(PACKAGE_VERSION_COMPATIBLE TRUE)
-    else()
-        set(PACKAGE_VERSION_COMPATIBLE FALSE)
-    endif()
+    set(PACKAGE_VERSION_COMPATIBLE TRUE)
+endif()
 
-    if(PACKAGE_FIND_VERSION STREQUAL PACKAGE_VERSION)
-        set(PACKAGE_VERSION_EXACT TRUE)
-    endif()
+if(PACKAGE_FIND_VERSION STREQUAL PACKAGE_VERSION)
+    set(PACKAGE_VERSION_EXACT TRUE)
 endif()

docs/cmake.md

@@ -19,12 +19,17 @@ That puts `kath` in `<prefix>/bin`, the `--lang` catalogues in
 ## Finding it
 
 ```cmake
-find_package(Booth 5.1 REQUIRED)
+find_package(Booth 0.5 REQUIRED)
 ```
 
 If the prefix is not one CMake already searches, point it there with
-`-DCMAKE_PREFIX_PATH=<prefix>`. Version matching is same-major: 5.1 is
-satisfied by any later 5.x, never by 6.x.
+`-DCMAKE_PREFIX_PATH=<prefix>`.
+
+While Booth is on 0.x the minor version has to match, so asking for 0.5 is
+satisfied by 0.5.2 but not by 0.6. That is deliberate: 0.x is where things
+are still allowed to move, and a package config that quietly accepted 0.6
+would be promising something Booth is not ready to promise. From 1.0 the
+usual same-major rule takes over.
 
 You get `Booth_VERSION`, `Booth_EXECUTABLE`, `Booth_LANG_DIR`, and a
 `Booth::kath` imported target.

examples/cmake/CMakeLists.txt

@@ -12,7 +12,7 @@
 cmake_minimum_required(VERSION 3.10)
 project(booth_cmake_example LANGUAGES NONE)
 
-find_package(Booth 5.1 REQUIRED)
+find_package(Booth 0.5 REQUIRED)
 message(STATUS "Booth ${Booth_VERSION} at ${Booth_EXECUTABLE}")
 
 booth_add_kernel(vadd_ptx    SOURCE vadd.cu BACKEND nvidia-ptx)

src/barracuda.h

@@ -5,11 +5,14 @@
 #include <stdio.h>
 #include <string.h>
 
-/* ---- Version ---- */
-#define BC_VERSION_MAJOR    5
-#define BC_VERSION_MINOR    1
-#define BC_VERSION_PATCH    0
-#define BC_VERSION_STRING   "5.01"
+/* ---- Version ----
+ * The 0.5 release was tagged v5.01, which was a typo for 0.5.1 and made the
+ * compiler look four major versions further along than it is. Corrected here;
+ * everything downstream derives from these three numbers. */
+#define BC_VERSION_MAJOR    0
+#define BC_VERSION_MINOR    5
+#define BC_VERSION_PATCH    2
+#define BC_VERSION_STRING   "0.5.2"
 
 /* The universe has limits. So do our buffers. No malloc, no madness. */
 #define BC_MAX_SOURCE       (4 * 1024 * 1024)