Merge pull request #227 from gn64/feature/vulkan_device_list

Add Vulkan GPU enumeration support (related to #160)

src/lib.rs

@@ -1,6 +1,9 @@
 #![allow(clippy::uninlined_format_args)]
 #![cfg_attr(test, feature(test))]
 
+#[cfg(feature = "vulkan")]
+pub mod vulkan;
+
 mod common_logging;
 mod error;
 mod ggml_logging_hook;

src/vulkan.rs

@@ -0,0 +1,72 @@
+use std::{ffi::CStr, os::raw::c_int};
+use whisper_rs_sys::{
+    ggml_backend_buffer_type_t, ggml_backend_vk_buffer_type, ggml_backend_vk_get_device_count,
+    ggml_backend_vk_get_device_description, ggml_backend_vk_get_device_memory,
+};
+
+#[derive(Debug, Clone)]
+pub struct VKVram {
+    pub free: usize,
+    pub total: usize,
+}
+
+/// Human-readable device information
+#[derive(Debug, Clone)]
+pub struct VkDeviceInfo {
+    pub id: i32,
+    pub name: String,
+    pub vram: VKVram,
+    /// Buffer type to pass to `whisper::Backend::create_buffer`
+    pub buf_type: ggml_backend_buffer_type_t,
+}
+/// Enumerate every physical GPU ggml can see.
+///
+/// Note: integrated GPUs are returned *after* discrete ones,
+/// mirroring ggml’s C logic.
+pub fn list_devices() -> Vec<VkDeviceInfo> {
+    unsafe {
+        let n = ggml_backend_vk_get_device_count();
+        (0..n)
+            .map(|id| {
+                // 256 bytes is plenty (spec says 128 is enough)
+                let mut tmp = [0i8; 256];
+                ggml_backend_vk_get_device_description(id as c_int, tmp.as_mut_ptr(), tmp.len());
+                let mut free = 0usize;
+                let mut total = 0usize;
+                ggml_backend_vk_get_device_memory(id, &mut free, &mut total);
+                VkDeviceInfo {
+                    id,
+                    name: CStr::from_ptr(tmp.as_ptr()).to_string_lossy().into_owned(),
+                    vram: VKVram { free, total },
+                    buf_type: ggml_backend_vk_buffer_type(id as usize),
+                }
+            })
+            .collect()
+    }
+}
+
+#[cfg(test)]
+mod vulkan_tests {
+    use super::*;
+
+    #[test]
+    fn enumerate_must_not_panic() {
+        let _ = list_devices();
+    }
+
+    #[test]
+    fn sane_device_info() {
+        let gpus = list_devices();
+        let mut seen = std::collections::HashSet::new();
+
+        for dev in &gpus {
+            assert!(seen.insert(dev.id), "duplicated id {}", dev.id);
+            assert!(!dev.name.trim().is_empty(), "GPU {} has empty name", dev.id);
+            assert!(
+                dev.vram.total >= dev.vram.free,
+                "GPU {} total < free",
+                dev.id
+            );
+        }
+    }
+}

sys/build.rs

@@ -119,10 +119,18 @@ fn main() {
         let _: u64 = std::fs::copy("src/bindings.rs", out.join("bindings.rs"))
             .expect("Failed to copy bindings.rs");
     } else {
-        let bindings = bindgen::Builder::default().header("wrapper.h");
+        let mut bindings = bindgen::Builder::default().header("wrapper.h");
 
         #[cfg(feature = "metal")]
-        let bindings = bindings.header("whisper.cpp/ggml/include/ggml-metal.h");
+        {
+            bindings = bindings.header("whisper.cpp/ggml/include/ggml-metal.h");
+        }
+        #[cfg(feature = "vulkan")]
+        {
+            bindings = bindings
+                .header("whisper.cpp/ggml/include/ggml-vulkan.h")
+                .clang_arg("-DGGML_USE_VULKAN=1");
+        }
 
         let bindings = bindings
             .clang_arg("-I./whisper.cpp/")

sys/wrapper.h

@@ -1,2 +1,6 @@
 #include <include/whisper.h>
 #include <ggml/include/ggml.h>
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif