Update to opencl3 0.9 crate

 - Update rust version to 1.56.1

 - Certain opencl3 APIs now required to be called as unsafe

Cargo.toml

@@ -3,7 +3,7 @@ name = "rust-gpu-tools"
 version = "0.6.1"
 authors = ["Keyvan Kambakhsh <[email protected]>", "porcuquine <[email protected]>"]
 description = "Rust OpenCL tools"
-edition = "2018"
+edition = "2021"
 homepage = "https://github.com/filecoin-project/rust-gpu-tools"
 license = "MIT/Apache-2.0"
 repository = "https://github.com/filecoin-project/rust-gpu-tools"
@@ -22,7 +22,8 @@ lazy_static = "1.2"
 log = "0.4.11"
 hex = "0.4.3"
 
-opencl3 = { version = "0.6", default-features = false, features = ["CL_VERSION_1_2"], optional = true }
+opencl3 = { version = "0.9", default-features = false, features = ["CL_VERSION_1_2"], optional = true }
+opencl-sys = "0.2.3"
 rustacuda = { package = "fil-rustacuda", version = "0.1.3", optional = true }
 once_cell = "1.8.0"
 temp-env = "0.2.0"

examples/add.rs

@@ -59,13 +59,13 @@ pub fn main() {
     let device = *Device::all().first().unwrap();
 
     // First we run it on CUDA.
-    let cuda_program = cuda(&device);
+    let cuda_program = cuda(device);
     let cuda_result = cuda_program.run(closures, ()).unwrap();
     assert_eq!(cuda_result, [6, 8, 10, 12]);
     println!("CUDA result: {:?}", cuda_result);
 
     // Then we run it on OpenCL.
-    let opencl_program = opencl(&device);
+    let opencl_program = opencl(device);
     let opencl_result = opencl_program.run(closures, ()).unwrap();
     assert_eq!(opencl_result, [6, 8, 10, 12]);
     println!("OpenCL result: {:?}", opencl_result);

rust-toolchain

@@ -1 +1 @@
-1.51.0
+1.56.1

src/cuda/mod.rs

@@ -114,7 +114,7 @@ impl Device {
 /// The majority of methods are the same as [`crate::opencl::Program`], so you can write code using this
 /// API, which will then work with OpenCL as well as CUDA kernels.
 // When compiled without the `opencl` feature, then the intra-doc link above will be broken.
-#[allow(broken_intra_doc_links)]
+#[allow(rustdoc::broken_intra_doc_links)]
 #[derive(Debug)]
 pub struct Program {
     context: rustacuda::context::UnownedContext,

src/device.rs

@@ -69,7 +69,7 @@ impl TryFrom<&str> for PciId {
 
     fn try_from(pci_id: &str) -> GPUResult<Self> {
         let mut bytes = [0; mem::size_of::<u16>()];
-        hex::decode_to_slice(pci_id.replace(":", ""), &mut bytes).map_err(|_| {
+        hex::decode_to_slice(pci_id.replace(':', ""), &mut bytes).map_err(|_| {
             GPUError::InvalidId(format!(
                 "Cannot parse PCI ID, expected hex-encoded string formated as aa:bb, got {0}.",
                 pci_id
@@ -111,7 +111,7 @@ impl TryFrom<&str> for DeviceUuid {
 
     fn try_from(uuid: &str) -> GPUResult<Self> {
         let mut bytes = [0; UUID_SIZE];
-        hex::decode_to_slice(uuid.replace("-", ""), &mut bytes)
+        hex::decode_to_slice(uuid.replace('-', ""), &mut bytes)
             .map_err(|_| {
                 GPUError::InvalidId(format!("Cannot parse UUID, expected hex-encoded string formated as aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee, got {0}.", uuid))
             })?;
@@ -137,7 +137,7 @@ impl fmt::Display for DeviceUuid {
 
 impl fmt::Debug for DeviceUuid {
     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-        write!(f, "{}", self.to_string())
+        write!(f, "{}", self)
     }
 }
 

src/opencl/mod.rs

@@ -121,7 +121,7 @@ impl Device {
 ///
 /// The majority of methods are the same as [`crate::cuda::Program`], so you can write code using this
 /// API, which will then work with OpenCL as well as CUDA kernels.
-#[allow(broken_intra_doc_links)]
+#[allow(rustdoc::broken_intra_doc_links)]
 pub struct Program {
     device_name: String,
     queue: CommandQueue,
@@ -157,7 +157,7 @@ impl Program {
                 "Building kernel ({}) from source: done.",
                 cached.to_string_lossy()
             );
-            let queue = CommandQueue::create(&context, context.default_device(), 0)?;
+            let queue = CommandQueue::create_default(&context, 0)?;
             let kernels = opencl3::kernel::create_program_kernels(&program)?;
             let kernels_by_name = kernels
                 .into_iter()
@@ -185,13 +185,13 @@ impl Program {
         debug!("Creating OpenCL program from binary.");
         let context = Context::from_device(&device.device)?;
         let bins = vec![&bin[..]];
-        let mut program =
-            opencl3::program::Program::create_from_binary(&context, context.devices(), &bins)?;
+        let mut program = unsafe {
+            opencl3::program::Program::create_from_binary(&context, context.devices(), &bins)
+        }?;
         if let Err(build_error) = program.build(context.devices(), "") {
             let log = program.get_build_log(context.devices()[0])?;
             return Err(GPUError::Opencl3(build_error, Some(log)));
-        }
-        let queue = CommandQueue::create(&context, context.default_device(), 0)?;
+        }        let queue = CommandQueue::create_default(&context, 0)?;
         let kernels = opencl3::kernel::create_program_kernels(&program)?;
         let kernels_by_name = kernels
             .into_iter()
@@ -248,21 +248,25 @@ impl Program {
         // The underlying buffer is `u8`, hence we need the number of bytes.
         let bytes_len = length * std::mem::size_of::<T>();
 
-        let mut buffer = opencl3::memory::Buffer::create(
-            &self.context,
-            CL_MEM_READ_WRITE,
-            bytes_len,
-            ptr::null_mut(),
-        )?;
+        let mut buffer = unsafe {
+            opencl3::memory::Buffer::create(
+                &self.context,
+                CL_MEM_READ_WRITE,
+                bytes_len,
+                ptr::null_mut(),
+            )?
+        };
         // Transmuting types is safe as long a sizes match.
         let bytes = unsafe {
             std::slice::from_raw_parts(slice.as_ptr() as *const T as *const u8, bytes_len)
         };
         // Write some data right-away. This makes a significant performance different.
-        self.queue
-            .enqueue_write_buffer(&mut buffer, CL_BLOCKING, 0, &[0u8], &[])?;
-        self.queue
-            .enqueue_write_buffer(&mut buffer, CL_BLOCKING, 0, &bytes, &[])?;
+        unsafe {
+            self.queue
+                .enqueue_write_buffer(&mut buffer, CL_BLOCKING, 0, &[0u8], &[])?;
+            self.queue
+                .enqueue_write_buffer(&mut buffer, CL_BLOCKING, 0, bytes, &[])?;
+        };
 
         Ok(Buffer::<T> {
             buffer,
@@ -287,7 +291,7 @@ impl Program {
             .kernels_by_name
             .get(name)
             .ok_or_else(|| GPUError::KernelNotFound(name.to_string()))?;
-        let mut builder = ExecuteKernel::new(&kernel);
+        let mut builder = ExecuteKernel::new(kernel);
         builder.set_global_work_size(global_work_size * local_work_size);
         builder.set_local_work_size(local_work_size);
         Ok(Kernel {
@@ -314,9 +318,10 @@ impl Program {
                 data.len() * std::mem::size_of::<T>(),
             )
         };
-        self.queue
-            .enqueue_write_buffer(&mut buffer.buffer, CL_BLOCKING, 0, &bytes, &[])?;
-
+        unsafe {
+            self.queue
+                .enqueue_write_buffer(&mut buffer.buffer, CL_BLOCKING, 0, bytes, &[])?;
+        }
         Ok(())
     }
 
@@ -325,15 +330,16 @@ impl Program {
         assert!(data.len() <= buffer.length, "Buffer is too small");
 
         // It is safe as long as the sizes match.
-        let mut bytes = unsafe {
+        let bytes = unsafe {
             std::slice::from_raw_parts_mut(
                 data.as_mut_ptr() as *mut T as *mut u8,
                 data.len() * std::mem::size_of::<T>(),
             )
         };
-        self.queue
-            .enqueue_read_buffer(&buffer.buffer, CL_BLOCKING, 0, &mut bytes, &[])?;
-
+        unsafe {
+            self.queue
+                .enqueue_read_buffer(&buffer.buffer, CL_BLOCKING, 0, bytes, &[])?;
+        };
         Ok(())
     }
 
@@ -362,27 +368,35 @@ pub trait KernelArgument {
 
 impl<T> KernelArgument for Buffer<T> {
     fn push(&self, kernel: &mut Kernel) {
-        kernel.builder.set_arg(&self.buffer);
+        unsafe {
+            kernel.builder.set_arg(&self.buffer);
+        }
     }
 }
 
 impl KernelArgument for i32 {
     fn push(&self, kernel: &mut Kernel) {
-        kernel.builder.set_arg(self);
+        unsafe {
+            kernel.builder.set_arg(self);
+        }
     }
 }
 
 impl KernelArgument for u32 {
     fn push(&self, kernel: &mut Kernel) {
-        kernel.builder.set_arg(self);
+        unsafe {
+            kernel.builder.set_arg(self);
+        }
     }
 }
 
 impl<T> KernelArgument for LocalBuffer<T> {
     fn push(&self, kernel: &mut Kernel) {
-        kernel
-            .builder
-            .set_arg_local_buffer(self.length * std::mem::size_of::<T>());
+        unsafe {
+            kernel
+                .builder
+                .set_arg_local_buffer(self.length * std::mem::size_of::<T>());
+        }
         kernel.num_local_buffers += 1;
     }
 }
@@ -431,7 +445,9 @@ impl<'a> Kernel<'a> {
                 "There cannot be more than one `LocalBuffer`.".to_string(),
             ));
         }
-        self.builder.enqueue_nd_range(&self.queue)?;
+        unsafe {
+            self.builder.enqueue_nd_range(self.queue)?;
+        }
         Ok(())
     }
 }

src/opencl/utils.rs

@@ -1,7 +1,7 @@
 use std::convert::TryFrom;
 
 use log::{debug, warn};
-use opencl3::device::CL_UUID_SIZE_KHR;
+use opencl_sys::CL_UUID_SIZE_KHR;
 use sha2::{Digest, Sha256};
 
 use crate::device::{DeviceUuid, PciId, Vendor};