gomlx · janpfeifer · Dec 19, 2024 · Dec 10, 2024 · Dec 10, 2024 · Dec 11, 2024
diff --git a/.github/workflows/go.yaml b/.github/workflows/go.yaml
@@ -42,7 +42,7 @@ jobs:
     - name: Install Go
       uses: actions/setup-go@v5
       with:
-        go-version: "1.22.x"
+        go-version: "1.23.x"
 
     - name: Install Gopjrt C library gomlx_xlabuilder and PJRT plugin
       shell: bash

diff --git a/README.md b/README.md
@@ -329,7 +329,7 @@ Also, see [this blog post](https://opensource.googleblog.com/2024/03/pjrt-plugin
   Because of https://github.com/golang/go/issues/13467 : C API's cannot be exported across packages, even within the same repo.
   Even a function as simple as `func Add(a, b C.int) C.int` in one package cannot be called from another. 
   So we need to wrap everything, and more than that, one cannot create separate sub-packages to handle separate concerns.
-  THis is also the reason the library `chelper.go` is copied in both `pjrt` and `xlabuilder` packages.
+  This is also the reason the library `chelper.go` is copied in both `pjrt` and `xlabuilder` packages.
 * **Why does PJRT spits out so much logging ? Can we disable it ?**
   This is a great question ... imagine if every library we use decided they also want to clutter our stderr?
   I have [an open question in Abseil about it](https://github.com/abseil/abseil-cpp/discussions/1700).
@@ -340,6 +340,18 @@ Also, see [this blog post](https://opensource.googleblog.com/2024/03/pjrt-plugin
   before calling `pjrt.GetPlugin`. But it may have unintended consequences, if some other library is depending
   on the fd 2 to work, or if a real exceptional situation needs to be reported and is not.
 
+## Environment Variables
+
+That help control or debug how **gopjrt** work:
+
+* `PJRT_PLUGIN_LIBRARY_PATH`: Path to search for PJRT plugins. **gopjrt** also searches in `/usr/local/lib/gomlx/pjrt`,
+  the standard library paths for the system and `$LD_LIBRARY_PATH`.
+* `XLA_DEBUG_OPTIONS`: If set, it is parsed as a `DebugOptions` proto that
+  is passed during the JIT-compilation (`Client.Compile()`) of a computation graph.
+  It is not documented how it works in PJRT (e.g. I observed a great slow down when this is set,
+  even if set to the default values), but [the proto has some documentation](https://github.com/gomlx/gopjrt/blob/main/protos/xla.proto#L40).
+* `GOPJRT_INSTALL_DIR` and `GOPJRT_NOSUDO`: used by the install scripts, see "Installing" section above.
+
 ## Links to documentation
 
 * [Google Drive Directory with Design Docs](https://drive.google.com/drive/folders/18M944-QQPk1E34qRyIjkqDRDnpMa3miN): Some links are outdated or redirected, but very valuable information.

diff --git a/c/WORKSPACE b/c/WORKSPACE
@@ -21,11 +21,12 @@ http_archive(
 # Notice bazel.sh scrape the line below for the OpenXLA version, the format
 # of the line should remain the same (the hash in between quotes), or bazel.sh
 # must be changed accordingly.
-OPENXLA_XLA_COMMIT_HASH = "90af2896ab4992ff14a1cd2a75ce02e43f46c090"  # From 2024-11-24
+# OPENXLA_XLA_COMMIT_HASH = "90af2896ab4992ff14a1cd2a75ce02e43f46c090"  # From 2024-11-24
+OPENXLA_XLA_COMMIT_HASH = "e2e8952ad0fac8833e9a78f9b3689e803ff8524f"  # From 2024-12-11
 
 http_archive(
     name = "xla",
-    sha256 = "a910124d546bc79edb685612edaa3d56153f0e0927f967e8defaf312b833d404",  # From 2024-11-24
+    sha256 = "5ec6919a25952fa790904983481ccb51ebbe20bbc53e15ddbb6d3e0b3aa3dfe1",  # From 2024-12-11
     strip_prefix = "xla-" + OPENXLA_XLA_COMMIT_HASH,
     urls = [
         "https://github.com/openxla/xla/archive/{hash}.zip".format(hash = OPENXLA_XLA_COMMIT_HASH),

diff --git a/chelper.go b/chelper.go
@@ -31,17 +31,15 @@ func cSizeOf[T any]() C.size_t {
 // It must be manually freed with cFree() by the user.
 func cMalloc[T any]() (ptr *T) {
 	size := cSizeOf[T]()
-	cPtr := (*T)(C.malloc(size))
-	C.memset(unsafe.Pointer(cPtr), 0, size)
+	cPtr := (*T)(C.calloc(1, size))
 	return cPtr
 }
 
 // cMallocArray allocates space to hold n copies of T in the C heap and initializes it to zero.
 // It must be manually freed with C.free() by the user.
 func cMallocArray[T any](n int) (ptr *T) {
-	size := cSizeOf[T]() * C.size_t(n)
-	cPtr := (*T)(C.malloc(size))
-	C.memset(unsafe.Pointer(cPtr), 0, size)
+	size := cSizeOf[T]()
+	cPtr := (*T)(C.calloc(C.size_t(n), size))
 	return cPtr
 }
 

diff --git a/cmd/run_coverage.sh b/cmd/run_coverage.sh
@@ -2,6 +2,6 @@
 
 # Run this from the root of gopjrt repository to generate docs/coverage.out with the coverage data.
 
-PACKAGE_COVERAGE="./pjrt ./xlabuilder"
-go test -v -cover -coverprofile docs/coverage.out -coverpkg ${PACKAGE_COVERAGE}
+PACKAGE_COVERAGE="github.com/gomlx/gopjrt/pjrt,github.com/gomlx/gopjrt/xlabuilder"
+go test -cover -coverprofile docs/coverage.out -coverpkg="${PACKAGE_COVERAGE}" ./... -test.count=1 -test.short
 go tool cover -func docs/coverage.out -o docs/coverage.out
diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.md
@@ -1,11 +1,31 @@
-# Next
+# v0.5.0 - 2024/12/19 - Adding direct access to PJRT buffers for CPU.
 
 * Added `install_linux_amd64_amazonlinux.sh` and pre-built libraries for amazonlinux (built using old glibc support).
 * Fixed installation scripts: s/sudo/$_SUDO. Also made them more verbose.
 * Removed dependency on `xargs` in installation script for Linux.
 * Improved documentation on Nvidia GPU card detection, and error message if not found. 
 * Updated GitHub action (`go.yaml`) to only change the README.md with the result of the change, if pushing to the
   `main` branch.
+* Added `prjt.arena` to avoid costly allocations for CGO calls, and merged some of CGO calls for general speed-ups.
+  The following functions had > 50% improvements on their fixed-cost (measured on transfers with 1 value, and minimal programs)
+  execution time (**not the variable part**): 
+  * `Buffer.ToHost()`
+  * `Client.BufferFromHost()`
+  * `LoadedExecutable.Execute()`
+* Added `BufferToHost` and `BufferFromHost` benchmarks.
+* Added support for environment variable `XLA_DEBUG_OPTIONS`: if set, it is parsed as a `DebugOptions` proto that
+  is passed to the JIT-compilation of a computation graph.
+* `LoadedExecutable.Execute()` now waits for the end of the execution (by setting
+  `PJRT_LoadedExecutable_Execute_Args.device_complete_events`).
+  Previous behavior lead to odd behavior and was undefined (not documented).
+* Package `dtypes`:
+  * Added tests;
+  * Added `SizeForDimensions()` to be used for dtypes that uses fractions of bytes (like 4 bits).
+* Added `Client.NewSharedBuffer` (and the lower level `client.CreateViewOfDeviceBuffer()`) to create buffers with shared
+  memory with the host, for faster input.
+  * Added `AlignedAlloc` and `AlignedFree` required by `client.CreateViewOfDeviceBuffer`.
+* Added `Buffer.Data` for direct access to a buffer's data. Undocumented in PJRT, and likely only works on CPU.
+* Fixed coverage script.
 
 # v0.4.9 - 2024-11-25